111 483 WHAT COMPUTERS CAN'T DO A CRITIQUE NEW YORK, EVANSTON, SAN FRANCISCO, LONDON 1817 WHAT COMPUTERS CAN'T DO OF ARTIFICIAL REASON By Hubert L. Dreyfus HARPER & ROW, PUBLISHERS The following materials are used with permission: The illustration on page 7 from Computers and Thought by Feigenbaum and Feldman. Copyright 1963 by Mcgraw-Hill, Inc. Used with permission of McGraw-Hill Book Com- pany. Quotations in Chapter 2 from Ross Quillian reprinted from "Semantic Memory" in Seman- tic Information Processing edited by Marvin Minsky by permission of The M.I.T. Press, Cambridge, Massachusetts. Copyright 1968 by the Massachusetts Institute of Technology. Quotations in Chapter 2 from Thomas G. Evans reprinted from "A Program for the Solution of a Class of Geometric- Analogy Intelligence Test Questions" in Semantic Infor- mation Processing edited by Marvin Minsky by permission of The M.I.T. Press, Cam- bridge, Massachusetts. Copyright 1968 by the Massachusetts Institute of Technology. Quotations in Chapter 7 from Anthony Oettinger reprinted from, and by permission of, American Documentation, July 1968, vol. 19, no. 3, pp. 295-298. Copyright 1968 by American Society for Information Science, 1 140 Connecticut Avenue, N.W., Washington, D.C 20036. WHAT COMPUTERS CAN'T DO: A CRITIQUE OF ARTIFICIAL REASON. Copyright 7972 by Hubert L. Dreyfus. All rights reserved. Printed in the United States of America. No part of this book may be used or reproduced in any manner whatsoever without written permission except in the case of brief quotations embodied in critical articles and reviews. For informa- tion address Harper & Row, Publishers, Inc., 49 East 33rd Street, New York, N. Y. 10016. Published simultaneously in Canada by Fitzhenry & Whiteside Limited, Toronto. FIRST EDITION STANDARD BOOK NUMBER: 06-011082-1 LIBRARY OF CONGRESS CATALOG CARD NUMBER: 67-22524 The difference between the mathematical mind {esprit de geome- trie) and the perceptive mind {esprit de finesse): the reason that math- ematicians are not perceptive is that they do not see what is before them, and that, accustomed to the exact and plain principles of mathematics, and not reasoning till they have well inspected and arranged their princi- ples, they are lost in matters of perception where the principles do not allow for such arrangement. . . . These principles are so fine and so numerous that a very delicate and very clear sense is needed to perceive them, and to judge rightly and justly when they are perceived, without for the most part being able to demonstrate them in order as in math- ematics; because the principles are not known to us in the same way, and because it would be an endless matter to undertake it. We must see the matter at once, at one glance, and not by a process of reasoning, at least to a certain degree. . . . Mathematicians wish to treat matters of percep- tion mathematically, and make themselves ridiculous . . . the mind . . . does it tacitly, naturally, and without technical rules. PASCAL, Pensees Contents Preface xi Acknowledgments xiv Introduction xv Part I. Ten Years of Research in Artificial Intelligence (1957-1967) 1. Phase I (1957-1962) Cognitive Simulation 3 i. Analysis of Work in Language Translation, Problem Solving, and Pattern Recognition II. The Underlying Significance of Failure to Achieve Predicted Results 2. Phase II (1962-1967) Semantic Information Processing 42 I. Analysis of Semantic Information Processing Programs II. Significance of Current Difficulties Conclusion 61 Part II. Assumptions Underlying Persistent Optimism Introduction 67 3. The Biological Assumption 71 4. The Psychological Assumption 75 ix I. Empirical Evidence for the Psychological Assumption: Critique of the Scientific Methodology of Cognitive Simulation II. A Priori Arguments for the Psychological Assumption 5. The Epistemological Assumption 101 I. A Mistaken Argument from the Success of Physics H. A Mistaken Argument from the Success of Modern Linguistics 6. The Ontological Assumption 118 Conclusion 137 Part III. Alternatives to the Traditional Assumptions Introduction 143 7. The Role of the Body in Intelligent Behavior 147 8. The Situation: Orderly Behavior Without Recourse to Rules 168 9. The Situation as a Function of Human Needs 184 Conclusion 193 CONCLUSION: The Scope and Limits of Artificial Reason The Limits of Artificial Intelligence 197 The Future of Artificial Intelligence Notes 221 Index 253 Preface In choosing to dissect artificial intelligence, Hubert Dreyfus has un- dertaken an inquiry of great public importance. This branch of science is seen by its practitioners as the basis for much more powerful versions of the computer technology that already pervades our society. As anyone can see who reads the daily press, many people are torn between hopes and fears aroused by digital computers, which they find mostly incom- prehensible and whose import therefore they cannot judge. But, as science lays claim to public support, so the public has a claim on critical analyses of science. Dreyfus serves all of us in venturing into an arcane technical field as a critical layman, a professional philosopher committed to questioning and analyzing the foundations of knowledge. Far, therefore, from shun- ning him as an unwelcome intruder or pelting him with acrimonious invective, artificial intelligence should welcome Dreyfus, draw on his correct insights, and set him right publicly, gently but ever so firmly, where misunderstandings or incomprehensions might flaw his logic. Dreyfus raises important and fundamental questions. One might there- fore expect the targets of his criticisms to react with greater human intelligence than when they simply shouted loud in response to his earlier sallies. The issues deserve serious public debate. They are too scientific to be left to philosophers and too philosophical to be left to scientists. Dreyfus sees agonizingly slow progress in all fundamental work on xi artificial intelligence. This he interprets as a sign of impenetrable barri- ers, rather than as the normal price for overcoming enormous technical and conceptual difficulties on the way to inevitable success. He sees artificial intelligence as limited by its assumption that the world is expli- cable in terms of elementary atomistic concepts, in a tradition traceable back to the Greeks. This insight challenges not only contemporary science and technology but also some of the foundations of Western philosophy. He puts in question the basic role that rules play in accepted ideas of what constitutes a satisfactory scientific explanation. Thereby he strikes at far more than the ability in principle of digital computers bound as they are to follow rules to exhibit intelligence of a kind which, according to his analysis, cannot be explained according to Kantian rules. He is too modern to ask his questions from a viewpoint that assumes that man and mind are somehow set apart from the physical universe and therefore not within reach of science. Quite to the contrary, he states explicitly his assumption that "there is no reason why, in principle, one could not construct an artificial embodied agent if one used components sufficiently like those which make up a human being." Instead, he points out that his questions are "philosophically interesting only if we restrict ourselves to asking if one can make such a robot by using a digital computer." Curiously enough to this technologist, Dreyfus's own philo- sophical arguments lead him to see digital computers as limited not so much by being mindless, as by having no body. This conclusion emerges from the contrast between the ability of human beings to "zero in" on relevant features of their environment while ignoring myriad irrelevancies and the enormous and admitted difficulty of artificial intelligence in determining what is relevant when the environment presented to a digital computer has not, in some way, been artificially constrained. The central statement of this theme is that "a person experiences the objects of the world as already interrelated and full of meaning. There is no justification for the assumption that we first experience isolated facts or snapshots of facts or momentary views of snapshots of isolated facts and then give them significance. This is the point that contemporary philosophers such as Heidegger and Wittgen- rrerace stein are trying to make/' The burden of artificial intelligence is indeed its apparent need to proceed in futility from the atom to the whole. People, on the other hand, effectively seem to perceive first a whole and only then, if necessary, analyze it into atoms. This, Dreyfus argues following Merleau-Ponty, is a consequence of our having bodies capable of an ongoing but unanalyzed mastery of their environment. Either Dreyfus's position or that of artificial intelligence might some day be corroborated or destroyed by new evidence from artificial intelli- gence itself, from psychology, neurophysiology, or other related disci- plines. Unless and until this happens, Dreyfus's work will stand for the layman as a lucid analysis of a difficult matter of great public moment. To the computer scientist concerned with progress in his specialty and with deeper understanding of the world, Dreyfus presents a profound challenge to the widespread idea that "knowledge consists of a large store of neutral data." Dreyfus clearly is not neutral. Anthony G. Oettinger Aiken Computation Laboratory Harvard University Acknowledgments The occasional acknowledgments scattered throughout the following pages only begin to reflect my indebtedness to a host of sympathetic and critical readers who not only weeded out mistakes but made many sub- stantive suggestions. Without the help of Ned Block, Susan Carey Block, Burton Bloom, Stuart Dreyfus, John Haugeland, Terrance Malick, An- thony Oettinger, Seymour Papert, George Rey, Charles Taylor, and Samuel Todes, this book would have been published much sooner and been easier to read, but also easier to dismiss. I am grateful to the American Council of Learned Societies, the Na- tional Science Foundation, and the RAND Corporation for supporting various stages of my research and writing, and to the Study Group on the Unity of Knowledge for enabling me to organize colloquia on several topics which subsequently found their way into the book. I also want to thank Rena Lieb for debugging an early version of the manuscript, and especially Barbara Behrendt, who deciphered the first draft and helped in more ways than I can mention. Introduction Since the Greeks invented logic and geometry, the idea that all reasoning might be reduced to some kind of calculation so that all arguments could be settled once and for all has fascinated most of the Western tradition's rigorous thinkers. Socrates was the first to give voice to this vision. The story of artificial intelligence might well begin around 450 B.C. when (according to Plato) Socrates demands of Euthyphro, a fellow Athenian who, in the name of piety, is about to turn in his own father for murder: "I want to know what is characteristic of piety which makes all actions pious . . . that I may have it to turn to, and to use as a standard whereby to judge your actions and those of other men. >M Socrates is asking Euthyphro for what modern computer theorists would call an "effective procedure," "a set of rules which tells us, from moment to moment, precisely how to behave.'* 2 Plato generalized this demand for moral certainty into an epistemolog- ical demand. According to Plato, all knowledge must be stateable in explicit definitions which anyone could apply. If one could not state his know-how in terms of such explicit instructions if his knowing how Notes begin on p. 221. [Citations are indicated by a superior figure. Substantive notes are indicated by a superior figure and an asterisk.] / XV Introduction / xvl could not be converted into knowing that it was not knowledge but mere belief. According to Plato, cooks, for example, who proceed by taste and intuition, and poets who work from inspiration, have no knowl- edge: what they do does not involve understanding and cannot be under- stood. More generally, what cannot be stated explicitly in precise instructions all areas of human thought which require skill, intuition, or a sense of tradition are relegated to some kind of arbitrary fum- bling. But Plato was not yet fully a cyberneticist (although according to Norbert Wiener he was the first to use the term), for Plato was looking for semantic rather than syntactic criteria. His rules presupposed that the person understood the meanings of the constitutive terms. In the Republic Plato says that Understanding (the rulelike level of his divided line representing all knowledge) depends on Reason, which involves a dialectical analysis and ultimately an intuition of the meaning of the fundamental concepts used in understanding. Thus Plato admits his instructions cannot be completely formalized. Similarly, a modern com- puter expert, Marvin Minsky, notes, after tentatively presenting a Pla- tonic notion of effective procedure: "This attempt at definition is subject to the criticism that the interpretation of the rules is left to depend on some person or agent." 3 Aristotle, who differed with Plato in this as in most questions concern- ing the application of theory to practice, noted with satisfaction that intuition was necessary to apply the Platonic rules: Yet it is not easy to find a formula by which we may determine how far and up to what point a man may go wrong before he incurs blame. But this difficulty of definition is inherent in every object of perception; such questions of degree are bound up with the circumstances of the individual case, where our only criterion is the perception. 4 For the Platonic project to reach fulfillment one breakthrough is required: all appeal to intuition and judgment must be eliminated. As Galileo discovered that one could find a pure formalism for describing physical motion by ignoring secondary qualities and teleological consid- erations, so, one might suppose, a Galileo of human behavior might Introduction / xvii succeed in reducing all semantic considerations (appeal to meanings) to the techniques of syntactic (formal) manipulation. The belief that such a total formalization of knowledge must be possi- ble soon came to dominate Western thought. It already expressed a basic moral and intellectual demand, and the success of physical science seemed to imply to sixteenth-century philosophers, as it still seems to suggest to thinkers such as Minsky, that the demand could be satisfied. Hobbes was the first to make explicit the syntactic conception of thought as calculation: "When a man reasons, he does nothing else but conceive a sum total from addition of parcels," he wrote, "for REASON ... is nothing but reckoning. . . ." 5 It only remained to work out the univocal parcels or "bits" with which this purely syntactic calculator could operate; Leibniz, the inventor of the binary system, dedicated himself to working out the necessary unam- biguous formal language. Leibniz thought he had found a universal and exact system of nota- tion, an algebra, a symbolic language, a "universal characteristic" by means of which "we can assign to every object its determined character- istic number." 6 In this way all concepts could be analyzed into a small number of original and undefined ideas; all knowledge could be ex- pressed and brought together in one deductive system. On the basis of these numbers and the rules for their combination all problems could be solved and all controversies ended: "if someone would (doubt my re- sults," Leibniz said, "I would say to him: 'Let us calculate, Sir,' and thus by taking pen and ink, we should settle the question." 7 Like a modern computer theorist announcing a program about to be written, Leibniz claims: Since, however, the wonderful interrelatedness of all things makes it extremely difficult to formulate explicitly the characteristic numbers of individual things, I have invented an elegant artifice by virtue of which certain relations may be represented and fixed numerically and which may thus then be further deter- mined in numerical calculation. 8 Nor was Leibniz reticent about the importance of his almost completed program. Introduction / xviii Once the characteristic numbers are established for most concepts, mankind will then possess a new instrument which will enhance the capabilities of the mind to far greater extent than optical instruments strengthen the eyes, and will supersede the microscope and telescope to the same extent that reason is superior to eyesight. 9 With this powerful new tool, the skills which Plato could not formal- ize, and so treated as confused thrashing around, could be recuperated as theory. In one of his "grant proposals" his explanations of how he could reduce all thought to the manipulation of numbers if he had money enough and time Leibniz remarks: the most important observations and turns of skill in all sorts of trades and professions are as yet unwritten. This fact is proved by experience when passing from theory to practice we desire to accomplish something. Of course, we can also write up this practice, since it is at bottom just another theory more complex and particular. . . . lo Leibniz had only promises, but in the work of George Boole, a math- ematician and logician working in the early nineteenth century, his program came one step nearer to reality. Like Hobbes, Boole supposed that reasoning was calculating, and he set out to "investigate the funda- mental laws of those operations of the mind by which reasoning is performed, to give expression to them in the symbolic language of a Calculus. . . ."" Boolean algebra is a binary algebra for representing elementary logical functions. If "a" and "fr" represent variables, "." represents "and," " + " represents "or," and "1" and "0" represent "true" and "false" respectively, then the rules governing logical manipulation can be writ- ten in algebraic form as follows: <2 -h a = a a + = a a -f 1 = 1 a- 0= a a- = a- 1 = a Western man was now ready to begin the calculation. Almost immediately, in the designs of Charles Babbage (1835), prac- tice began to catch up to theory. Babbage designed what he called an "Analytic Engine" which, though never built, was to function exactly like a modern digital computer, using punched cards, combining logical Introduction / xix and arithmetic operations, and making logical decisions along the way based upon the results of its previous computations. An important feature of Babbage's machine was that it was digital. There are two fundamental types of computing machines: analogue and digital. Analogue computers do not compute in the strict sense of the word. They operate by measuring the magnitude of physical quantities. Using physical quantities, such as voltage, duration, angle of rotation of a disk, and so forth, proportional to the quantity to be manipulated, they combine these quantities in a physical way and measure the result. A slide rule is a typical analogue computer. A digital computer as the word digit, Latin for "finger," implies represents all quantities by dis- crete states, for example, relays which are open or closed, a dial which can assume any one often positions, and so on, and then literally counts in order to get its result. Thus, whereas analogue computers operate with continuous quanti- ties, all digital computers are discrete state machines. As A. M. Turing, famous for defining the essence of a digital computer, puts it: [Discrete state machines] move by sudden jumps or clicks from one quite definite state to another. These states are sufficiently different for the possibility of confusion between them to be ignored. Strictly speaking there are no such machines. Everything really moves continuously. But there are many kinds of machines which can profitably be thought of as being discrete state machines. For instance in considering the switches for a lighting system it is a convenient fiction that each switch must be definitely on or definitely off. There must be intermedi- ate positions, but for most purposes we can forget about them. 12 Babbage's ideas were too advanced for the technology of his time, for there was no quick efficient way to represent and manipulate the digits. He had to use awkward mechanical means, such as the position of cogwheels, to represent the discrete states. Electric switches, however, provided the necessary technological breakthrough. When, in 1944, H. H. Aiken actually built the first practical digital computer, it was elec- tromechanical using about 3000 telephone relays. These were still slow, however, and it was only with the next generation of computers using vacuum tubes that the modern electronic computer was ready. Ready for anything. For, since a digital computer operates with ab- Introduction / xx stract symbols which can stand for anything, and logical operations which can relate anything to anything, any digital computer (unlike an analogue computer) is a universal machine. First, as Turing puts it, it can simulate any other digital computer. This special property of digital computers, that they can mimic any discrete state machine, is described by saying that they are universal machines. The existence of machines with this property has the important consequence that, considera- tions of speed apart, it is unnecessary to design various new machines to do various computing processes. They can all be done with one digital computer, suitably programmed for each case. It will be seen that as a consequence of this all digital computers are in a sense equivalent. 13 Second, and philosophically more significant, any process which can be formalized so that it can be represented as series of instructions for the manipulation of discrete elements, can, at least in principle, be repro- duced by such a machine. Thus even an analogue computer, provided that the relation of its input to its output can be described by a precise mathematical function, can be simulated on a digital machine. 14 * But such machines might have remained overgrown adding machines, had not Plato's vision, refined by two thousand years of metaphysics, found in them its fulfillment. At last here was a machine which operated according to syntactic rules, on bits of data. Moreover, the rules were built into the circuits of the machine. Once the machine was pro- grammed there was no need for interpretation; no appeal to human intuition and judgment. This was just what Hobbes and Leibniz had ordered, and Martin Heidegger appropriately saw in cybernetics the culmination of the philosophical tradition. 15 * Thus while practical men like Eckert and Mauchly, at the University of Pennsylvania, were designing the first electronic digital machine, theo- rists, such as Turing, trying to understand the essence and capacity of such machines, became interested in an area which had thus far been the province of philosophers: the nature of reason itself. In 1950, Turing wrote an influential article, "Computing Machinery and Intelligence," in which he points out that "the present interest in 'thinking machines' has been aroused by a particular kind of machine, usually called an 'electronic computer' or a 'digital computer.' " 16 He Introduction / xxi then takes up the question "Can [such] machines think?" To decide this question Turing proposes a test which he calls the imitation game: The new form of the problem can be described in terms of a game which we call the "imitation game." It is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game he says either "X is A and Y is B" or "X is B and Y is A." The interrogator is allowed to put questions to A and B thus: C: Will X please tell me the length of his or her hair? Now suppose X is actually A, then A must answer. It is A's object in the game to try to cause C to make the wrong identification. His answer might therefore be "My hair is shingled, and the longest strands are about nine inches long." In order that tones of voice may not help the interrogator the answers should be written, or better still, typewritten. The ideal arrangement is to have a tele- printer communicating between the two rooms. Alternatively, the question and answers can be repeated by an intermediary. The object of the game for the third player (B) is to help the interrogator. The best strategy for her is probably to give truthful answers. She can add such things as "I am the woman, don't listen to him!" to her answers, but it will avail nothing as the man can make similar remarks. We now ask the question, "What will happen when a machine takes the part of A in this game?" Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, "Can machines think?" 17 This test has become known as the Turing Test. Philosophers may doubt whether merely behavioral similarity could ever give adequate ground for the attribution of intelligence, 18 but as a goal for those actually trying to construct thinking machines, and as a criterion for critics to use in evaluating their work, Turing's test was just what was needed. Of course, no digital computer immediately volunteered or was drafted for Turing's game. In spite of its speed, accuracy, and universal- ity, the digital computer was still nothing more than a general-symbol manipulating device. The chips, however, were now down on the old Leibnizian bet. The time was ripe to produce the appropriate symbolism Introduction / xxii and the detailed instructions by means of which the rules of reason could be incorporated in a computer program. Turing had grasped the possibil- ity and provided the criterion for success, but his article ended with only the sketchiest suggestions about what to do next: We may hope that machines will eventually compete with men in all purely intellectual fields. But which are the best ones to start with? Even this is a difficult decision. Many people think that a very abstract activity, like the playing of chess, would be best. It can also be maintained that it is best to provide the machine with the best sense organs that money can buy, and then teach it to understand and speak English. This process could follow the normal teaching of a child. Things would be pointed out and named, etc. Again I do not know what the right answer is, but I think both approaches should be tried. 19 A technique was still needed for finding the rules which thinkers from Plato to Turing assumed must exist a technique for converting any practical activity such as playing chess or learning a language into the set of instructions Leibniz called a theory. Immediately, as if following Turing's hints, work got under way on chess and language. The same year Turing wrote his article, Claude E. Shannon, the inventor of infor- mation theory, wrote an article on chess-playing machines in which he discussed the options facing someone trying to program a digital com- puter to play chess. Investigating one particular line of play for 40 moves would be as bad as investi- gating all lines for just two moves. A suitable compromise would be to examine only the important possible variations that is, forcing moves, captures and main threats and cany out the investigation of the possible moves far enough to make the consequences of each fairly clear. It is possible to set up some rough criteria for selecting important variations, not as efficiently as a chess master, but sufficiently well to reduce the number of variations appreciably and thereby permit a deeper investigation of the moves actually considered. 20 Shannon did not write a chess program, but he believed that "an elec- tronic computer programmed in this manner would play a fairly strong game at speeds comparable to human speeds." 21 In 1955 Allen Newell wrote a sober survey of the problems posed by the game of chess and suggestions as to how they might be met. Newell notes that "These [suggested] mechanisms are so complicated that it is Introduction / xxm impossible to predict whether they will work." 22 The next year, however, brought startling success. A group at Los Alamos produced a program which played poor but legal chess on a reduced board. In a review of this work, Allen Newell, J. C. Shaw, and H. A. Simon concluded: "With very little in the way of complexity, we have at least entered the arena of human play we can beat a beginner." 23 And by 1957, Alex Bernstein had a program for the IBM 704 which played two "passable amateur games." 24 Meanwhile, Anthony Oettinger was working on the other Turing line. Having already in 1952 programmed a machine which simulated simple conditioning, increasing or decreasing a set response on the basis of positive or negative reinforcement, Oettinger turned to the problem of language translation and programmed a Russian-English mechanical dictionary. Further research in these directions, it seemed, might lead to a computer which could be taught to associate words and objects. But neither of these approaches offered anything like a general theory of intelligent behavior. What was needed were rules for converting any sort of intelligent activity into a set of instructions. At this point Herbert Simon and Allen Newell, analyzing the way a student proceeded to solve logic problems, noted that their subjects tended to use rules or shortcuts which were not universally correct, but which often helped, even if they sometimes failed. Such a rule of thumb might be, for example: always try to substitute a shorter expression for a longer one. Simon and Newell decided to try to simulate this practical intelligence. The term "heuristic program" was used to distinguish the resulting programs from programs which are guaranteed to work, so-called algorithmic programs which follow an exhaustive method to arrive at a solution, but which rapidly become unwieldy when dealing with practical problems. This notion of a rule of practice provided a breakthrough for those looking for a way to program computers to exhibit general problem- solving behavior. Something of the excitement of this new idea vibrates in the first paragraph of Newell, Shaw, and Simon's classic article "Em- pirical Explorations with the Logic Theory Machine: A Case Study in Heuristics." Introduction / xxiv This is a case study in problem-solving, representing part of a program of research on complex information-processing systems. We have specified a system for finding proofs of theorems in elementary symbolic logic, and by programming a computer to these specifications, have obtained empirical data on the problem- solving process in elementary logic. The program is called the Logic Theory Machine (LT); it was devised to learn how it is possible to solve difficult problems such as proving mathematical theorems, discovering scientific laws from data, playing chess, or understanding the meaning of English prose. The research reported here is aimed at understanding the complex processes (heuristics) that are effective in problem-solving. Hence, we are not interested in methods that guarantee solutions, but which require vast amounts of computa- tion. Rather, we wish to understand how a mathematician, for example, is able to prove a theorem even though he does not know when he starts how, or if, he is going to succeed. 25 But Newell and Simon soon realized that even this approach was not general enough. The following year (1957) they sought to abstract the heuristics used in the logic marine, and apply them to a range of similar problems. This gave rise to a program called the General Problem Solver or GPS. The motivation and orientation of the work on the General Problem Solver are explained in Newell, Shaw, and Simon's first major report on the enterprise. This paper ... is part of an investigation into the extremely complex processes that are involved in intelligent, adaptive, and creative behavior. . . . Many kinds of information can aid in solving problems: information may suggest the order in which possible solutions should be examined; it may rule out a whole class of solutions previously thought possible; it may provide a cheap test to distinguish likely from unlikely possibilities; and so on. All these kinds of information are heuristics things that aid discovery. Heuristics seldom pro- vide infallible guidance. . . . Often they "work," but the results are variable and success is seldom guaranteed. 26 To convey a sense of the general heuristics their program employed, Newell and Simon introduced an example of everyday intelligent be- havior: I want to take my son to nursery school. What's the difference between what I have and what I want? One of distance. What changes distance? My automobile. My automobile won*t work. What's needed to make it work? A new battery. What has new batteries? An auto repair shop. I want the repair shop to put in Introduction / xxv a new battery; but the shop doesn't know I need one. What is the difficulty? One of communication. What allows communication? A telephone. . . . And so on. This kind of analysis classifying things in terms of the functions they serve, and oscillating among ends, functions required, and means that perform them forms the basic system of heuristic of GPS. More precisely, this means-end system of heuristic assumes the following: 1 . If an object is given that is not the desired one, differences will be detectable between the available object and the desired object. 2. Operators affect some features of their operands and leave others un- changed. Hence operators can be characterized by the changes they produce and can be used to try to eliminate differences between the objects to which they are applied and desired objects. 3. Some differences will prove more difficult to affect than others. It is profita- ble, therefore, to try to eliminate "difficult" differences, even at the cost of introducing new differences of lesser difficulty. This process can be repeated as long as progress is being made toward eliminating the more difficult differences. 27 With digital computers solving such problems as how to get three cannibals and three missionaries across a river without the cannibals eating the missionaries, it seemed that finally philosophical ambition had found the necessary technology: that the universal, high-speed computer had been given the rules for converting reasoning into reckoning. Simon and Newell sensed the importance of the moment and jubilantly an- nounced that the era of intelligent machines was at hand. We have begun to learn how to use computers to solve problems, where we do not have systematic and efficient computational algorithms. And we now know, at least in a limited area, not only how to program computers to perform such problem-solving activities successfully; we know also how to program computers to learn to do these things. In short, we now have the elements of a theory of heuristic (as contrasted with algorithmic) problem solving; and we can use this theory both to understand human heuristic processes and to simulate such processes with digital computers. Intuition, insight, and learning are no longer exclusive possessions of humans: any large high-speed computer can be programmed to exhibit them also. 28 This field of research, dedicated to using digital computers to simulate intelligent behavior, soon came to be known as "artificial intelligence." One should not be misled by the name. No doubt an artificial nervous system sufficiently like the human one, with other features such as sense Introduction / xxvi organs and a body, would be intelligent. But the term "artificial" does not mean that workers in artificial intelligence are trying to build an artificial man. Given the present state of physics, chemistry, and neuro- physiology, such an undertaking is not feasible. Simon and the pioneers of artificial intelligence propose to produce something more limited: a heuristic program which will enable a digital information-processing machine to exhibit intelligence. Likewise, the term "intelligence" can be misleading. No one expects the resulting robot to reproduce everything that counts as intelligent behavior in human beings. It need not, for example, be able to pick a good wife, or get across a busy street. It must only compete in the more objective and disembodied areas of human behavior, so as to be able to win at Turing's game. This limited objective of workers in artificial intelligence is just what gives such work its overwhelming significance. These last metaphysicians are staking everything on man's ability to formalize his behavior; to bypass brain and body, and arrive, all the more surely, at the essence of rationality. Computers have already brought about a technological revolution comparable to the Industrial Revolution. If Simon is right about the imminence of artificial intelligence, they are on the verge of creating an even greater conceptual revolution a change in our understanding of man. Everyone senses the importance of this revolution, but we are so near the events that it is difficult to discern their significance. This much, however, is clear. Aristotle defined man as a rational animal, and since then reason has been held to be of the essence of man. If we are on the threshold of creating artificial intelligence we are about to see the tri- umph of a very special conception of reason. Indeed, if reason can be programmed into a computer, this will confirm an understanding of the nature of man, which Western thinkers have been groping toward for two thousand years but which they only now have the tools to express and implement. The incarnation of this intuition will drastically change our understanding of ourselves. If, on the other hand, artificial intelli- gence should turn out to be impossible, then we will have to distinguish Introduction / xxvii human from artificial reason, and this too will radically change our view of ourselves. Thus the moment has come either to face the truth of the tradition's deepest intuition or to abandon what has passed for an under- standing of man's nature for two thousand years. Although it is perhaps too early for a full answer, we must make an attempt to determine the scope and limits of the sort of reason which has come fully into force since the perfection of the "analytical engine." We must try to understand to what extent artificial intelligence is possible, and if there are limits to the possibility of computer simulation of intelli- gent behavior, we must determine those limits and their significance. What we learn about the limits of intelligence in computers will tell us something about the character and extent of human intelligence. What is required is nothing less than a critique of artificial reason. II The need for a critique of artificial reason is a special case of a general need for critical caution in the behavioral sciences. Chomsky remarks that in these sciences "there has been a natural but unfortunate tendency to 'extrapolate,' from the thimbleful of knowledge that has been attained in careful experimental work and rigorous data-processing, to issues of much wider significance and of great social concern." He concludes that the experts have the responsibility of making clear the actual limits of their understanding and of the results they have so far achieved. A careful analysis of these limits will demonstrate that in virtually every domain of the social and behavioral sciences the results achieved to date will not support such "extrapola- tion." 29 Artificial intelligence, at first glance, seems to be a happy exception to this pessimistic principle. Every day we read that digital computers play chess, translate languages, recognize patterns, and will soon be able to take over our jobs. In fact this now seems like child's play. Literally! In a North American Newspaper Alliance release, dated December 1968, entitled "A Computer for Kids" we are told that Introduction / xxviii Cosmos, the West German publishing house . . . has come up with a new idea in gifts. . . . It's a genuine (if small) computer, and it costs around $20. Battery operated, it looks like a portable typewriter. But it can be programmed like any big computer to translate foreign languages, diagnose illnesses, even provide a weather forecast. And in a recent Life Magazine article (Nov. 20, 1970) entitled "Meet Shakey, The First Electronic Person," the wide-eyed reader is told of a computer "made up of five major systems of circuitry that correspond quite closely to basic human faculties sensation, reason, language, memory [and] ego." According to the article, this computer "sees," "understands," "learns," and, in general, has "demonstrated that ma- chines can think." Several distinguished computer scientists are quoted as predicting that in from three to fifteen years "we will have a machine with the general intelligence of an average human being . . . and in a few months [thereafter] it will be at genius level. . . ." The complete robot may be a few years off, of course, but anyone interested in the prospective situation at the turn of the century can see in the film 2001: A Space Odyssey a robot named HAL who is cool, conversational, and very nearly omniscient and omnipotent. And this film is not simply science-fiction fantasy. A Space Odyssey was made with scrupulous documentation. The director, Stanley Kubrick, consulted the foremost computer specialists so as not to be misled as to what was at least remotely possible. Turing himself had in 1950 affirmed his belief that "at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted." 30 And the technical consultant for the film, Professor Marvin Minsky, working on an early prototype of HAL in his laboratory at M.I.T., assured Kubrick that Turing was, if anything, too pessimistic. That Minsky was not misunderstood by Kubrick is clear from Min- sky's editorial for Science Journal which reads like the scenario for 2007: At first machines had simple claws. Soon they will have fantastically graceful articulations. Computers' eyes once could sense only a hole in a card. Now they recognize shapes on simple backgrounds. Soon they will rival man's analysis of Introduction / X xix his environment. Computer programs once merely added columns of figures. Now they play games well, understand simple conversations, weigh many factors in decisions. What next? Today, machines solve problems mainly according to the principles we build into them. Before long, we may learn how to set them to work upon the very special problem of improving their own capacity to solve problems. Once a certain threshold is passed, this could lead to a spiral of acceleration and it may be hard to perfect a reliable 'governor' to restrain it. 31 It seems that there may be no limit to the range and brilliance of the properly programmed computer. It is no wonder that among philoso- phers of science one finds an assumption that machines can do every- thing people can do, followed by an attempt to interpret what this bodes for the philosophy of mind; while among moralists and theologians one finds a last-ditch retrenchment to such highly sophisticated behavior as moral choice, love, and creative discovery, claimed to be beyond the scope of any machine. Thinkers in both camps have failed to ask the preliminary question whether machines can in fact exhibit even elemen- tary skills like playing games, solving simple problems, reading simple sentences and recognizing patterns, presumably because they are under the impression, fostered by the press and artificial-intelligence research- ers such as Minsky, that the simple tasks and even some of the most difficult ones have already been or are about to be accomplished. To begin with, then, these claims must be examined. It is fitting to begin with a prediction made by Herbert Simon in 1957 as his General Problem Solver seemed to be opening up the era of artificial intelligence: It is not my aim to surprise or shock you. . . . But the simplest way I can summarize is to say that there are now in the world machines that think, that learn and that create. Moreover, their ability to do these things is going to increase rapidly until in a visible future the range of problems they can han- dle will be coextensive with the range to which the human mind has been ap- plied. Simon then predicts, among other things, 1. That within ten years a digital computer will be the world's chess champion, unless the rules bar it from competition. Introduction / xxx 2. That within ten years a digital computer will discover and prove an impor- tant new mathematical theorem. 3. That within ten years most theories in psychology will take the form of computer programs, or of qualitative statements about the characteristics of computer programs. 32 Unfortunately, the tenth anniversary of this historic talk went unno- ticed, and workers in artificial intelligence did not, at any of their many national and international meetings, take time out from their progress reports to confront these predictions with the actual achievements. Now fourteen years have passed, and we are being warned that it may soon be difficult to control our robots. It is certainly high time to measure this original prophecy against reality. Already in the five years following Simon's predictions, publications suggested that the first of Simon's forecasts had been half-realized, and that considerable progress had been made in fulfilling his second predic- tion. This latter, the theorem-discovery prediction, was "fulfilled" by W. R. Ashby (one of the leading authorities in the field) when, in a review of Feigenbaum and Feldman's anthology Computers and Thought, he hailed the mathematical power of the properly programmed computer: "Gelernter's theorem-proving program has discovered a new proof of the pons asinorum that demands no construction." This proof, Dr. Ashby goes on to say, is one which "the greatest mathematicians of 2000 years have failed to notice . ." . which would have evoked the highest praise had it occurred." 33 The theorem sounds important, and the naive reader cannot help sharing Ashby's enthusiasm. A little research, however, reveals that the pons asinorum, or ass's bridge, is the elementary theorem proved in Euclidian geometry namely that the opposite angles of an isosceles triangle are equal. Moreover, the first announcement of the "new" proof "discovered" by the machine is attributed to Pappus (A.D. 300). 34 There is a striking disparity between Ashby's excitement and the antiquity and simplicity of this proof. We are still a long way from "the important mathematical theorem" to be found by 1967. The chess-playing story is more involved and might serve as a model for a study of the production of intellectual smog in this area. In 1958, Introduction / xxxi the year after Simon's prediction, Newell, Shaw, and Simon presented an elaborate chess-playing program. As described in their classic paper, "Chess-Playing Programs and the Problem of Complexity," their pro- gram was "not yet fully debugged,*' so that one "cannot say very much about the behavior of the program." 35 Still, it is clearly "good in [the] . . . opening." 36 This is the last detailed published report on the program. In the same year, however, Newell, Shaw, and Simon announced: "We have written a program that plays chess," 37 and Simon, on the basis of this success, revised his earlier prediction: In another place, we have predicted that within ten years a computer will discover and prove an important mathematical theorem. On the basis of our experience with the heuristics of logic and chess, we are willing to add the further prediction that only moderate extrapolation is required from the capacities of programs already in existence to achieve the additional problem-solving power needed for such simulation. 38 Public gullibility and Simon's enthusiasm was such that Newell, Shaw, and Simon's claims concerning their still bugged program were sufficient to launch the chess machine into the realm of scientific mythology. In 1959, Norbert Wiener, escalating the claim that the program was "good in the opening," informed the NYU Institute of Philosophy that "chess- playing machines as of now will counter the moves of a master game with the moves recognized as right in the text books, up to some point in the middle game." 39 In the same symposium, Michael Scriven moved from the ambiguous claim that "machines now play chess" to the positive assertion that "machines are already capable of a good game." 40 In fact, in its few recorded games, the Newell, Shaw, Simon program played poor but legal chess, and in its last official bout (October 1960) was beaten in 35 moves by a ten-year-old novice. Fact, however, had ceased to be relevant. While their program was losing its five or six poor games and the myth they had created was holding its own against masters in the middle game Newell, Shaw, and Simon kept silent. When they speak again, three years later, they do not report their difficulties and disappointment. Rather, as if to take up where the myth left off, Simon published an article in Behavioral Science announcing a program which played Introduction / xxxii "highly creative" chess end games involving "combinations as difficult as any that have been recorded in chess history." 41 That the program restricts these end games to dependence on continuing checks, so that the number of relevant moves is greatly reduced, is mentioned but not emphasized. On the contrary, it is misleadingly implied that similar simple heuristics would account for master play even in the middle game/ 2 * Thus, the article gives the impression that the chess prediction is almost realized. With such progress, the chess championship may be claimed at any moment. Indeed, a Russian cyberneticist, upon hearing of Simon's ten-year estimate, called it "conservative." 43 And Fred Gruenberger at RAND suggested that a world champion is not enough that we should aim for "a program which plays better than any man could." 44 This regenerating confusion makes one think of the mythical French beast which is supposed to secrete the fog necessary for its own respiration. Reality comes limping along behind these impressive pronounce- ments. Embarrassed by my expose of the disparity between their enthusi- asm and their results, AI workers finally produced a reasonably competent program. R. Greenblatt's program called MacHack did in fact beat the author, 45 * a rank amateur, and has been entered in several tournaments in which it won a few games. This limited success revived hopes and claims. Seymour Papert, the second in command at the M.I.T. robot project, leaped in to defend Simon's prediction, asserting that "as a statement of what researchers in the field consider to be a possible goal for the near future, this is a reasonable statement." 46 And on page 1 of the October 1968 issue of Science Journal, Donald Michie, the leader of England's artificial intelligentsia, writes that "today machines can play chess at championship level." 47 However, chess master de Groot, discuss- ing the earlier chess programs, once said: "programs are still very poor chess players and I do not have much hope for substantial improvement in the future." And another chess master, Eliot Hearst, discussing the M.I.T. program in Psychology Today, adds: "De Groot's comment was made in 1964 and MacHack's recent tournament showing would not require him to revise his opinion." 48 Nor would most recent events. Greenblatt's program has been gradually improved, but it seems to have Introduction / xxxiii reached a point of saturation. During the past two years, it lost all games in the tournaments in which it had been entered, and received no further publicity. We shall soon see that given the limitations of digital comput- ers this is just what one would expect. It is to Greenblatt's credit that even in the heyday of MacHack he made no prediction; as for Simon and the world championship, the ten years are well up, and the computer is at best a class C amateur. 49 * This rapid rundown of the state of the art vis-a-vis two of Simon's predictions has, I hope, cleared the air. It is essential to be aware at the outset that despite predictions, press releases, films, and warnings, artifi- cial intelligence is a promise and not an accomplished fact. Only then can we begin our examination of the actual state and future hopes of artificial intelligence at a sufficiently rudimentary level. The field of artificial intelligence has many divisions and subdivisions, but the most important work can be classified into four areas: game playing, language translating, problem solving, and pattern recognition. We have already discussed the state of game-playing research. We shall now look at the work in the remaining three fields in detail. In Part I my general thesis will be that the field of artificial intelligence exhibits a recurring pattern: early, dramatic success followed by sudden unex- pected difficulties. This pattern occurs in all four areas, in two phases each lasting roughly five years. The work from 1957 to 1962 (Chapter 1), is concerned primarily with Cognitive Simulation (fcS) the use of heuristic programs to simulate human behavior by attempting to re- produce the steps by which human beings actually proceed. The second period (Chapter 2) is predominantly devoted to semantic information processing. This is artificial intelligence in a narrower sense than I have been using the term thus far. AI (for this restricted sense I shall use the initials) is the attempt to simulate human intelligent behavior using programming techniques which need bear little or no resemblance to human mental processes. The difficulties confronting this approach have just begun to emerge. The task of the rest of Part I is to discover the underlying common source of all these seemingly unconnected setbacks. These empirical difficulties, these failures to achieve predicted prog- Introduction / xxxiv ress, never, however, discourage the researchers, whose optimism seems to grow with each disappointment. We therefore have to ask what as- sumptions underlie this persistent optimism in the face of repeated fail- ures. Part II attempts to bring to light four deeply rooted assumptions or prejudices which mask the gravity of the current impasse, and to lay bare the conceptual confusion to which these prejudices give rise. But these prejudices are so deeply rooted in our thinking that the only alternative to them seems to be an obscurantist rejection of the possibility of a science of human behavior. Part III attempts to answer this objec- tion, insofar as it can be answered, by presenting an alternative to these traditional assumptions, drawing on the ideas of twentieth-century thinkers whose work is an implicit critique of artificial reason, although it has not before been read in this light. We shall then, in the Conclusion, be in a position to characterize artificial reason and indicate its scope and limits. This in turn will enable us to distinguish among various forms of intelligent behavior and to judge to what extent each of these types of intelligent behavior is pro- grammable in practice and in principle. If the order of argument presented above and the tone of my opening remarks seem strangely polemical for an effort in philosophical analysis, I can only point out that, as we have already seen, artificial intelligence is a field in which the rhetorical presentation of results often substitutes for research, so that research papers resemble more a debater's brief than a scientific report. Such persuasive marshaling of facts can only be an- swered in kind. Thus the accusatory tone of Part I. In Part II, however, I have tried to remain as objective as possible in testing fundamental assumptions, although I know from experience that challenging these assumptions will produce reactions similar to those of an insecure be- liever when his faith is challenged. For example, the year following the publication of my first investiga- tion of work in artificial intelligence, the RAND Corporation held a meeting of experts in computer science to discuss, among other topics, my report. Only an "expurgated" transcript of this meeting has been released to the public, but even there the tone of paranoia which per- vaded the discussion is present on almost every page. My report is called Introduction / xxxv "sinister," "dishonest," "hilariously funny," and an "incredible misrep- resentation of history." When, at one point, Dr. J. C. R. Licklider, then of IBM, tried to come to the defense of my conclusion that work should be done on man-machine cooperation, Seymour Papert of M.I.T. re- sponded: I protest vehemently against crediting Dreyfus with any good. To state that you can associate yourself with one of his conclusions is unprincipled. Dreyfus* concept of coupling men with machines is based on thorough misunderstanding of the problems and has nothing in common with any good statement that might go by the same words. 50 The causes of this panic-reaction should themselves be investigated, but that is a job for psychology, or the sociology of knowledge. However, in anticipation of the impending outrage I want to make absolutely clear from the outset that what I am criticizing is the implicit and explicit philosophical assumptions of Simon and Minsky and their co-workers, not their technical work. True, their philosophical prejudices and naivete distort their own evaluation of their results, but this in no way detracts from the importance and value of their research on specific techniques such as list structures, and on more general problems such as data-base organization and access, compatibility theorems, and so forth. The fun- damental ideas that they have contributed in these areas have not only made possible the limited achievements in artificial intelligence but have contributed to other more flourishing areas of computer science. In some restricted ways even AI can have, and presumably will have practical value despite what I shall try to show are its fundamental limitations. (I restrict myself to AI because it is not clear that naive Cognitive Simulation, as it is now practiced, can have any value at all, except perhaps as a striking demonstration of the fact that in behaving intelligently people do not process information like a heuristically pro- grammed digital computer.) An artifact could replace men in some tasks for example, those involved in exploring planets without performing the way human beings would and without exhibiting human flexibility. Research in this area is not wasted or foolish, although a balanced view of what can and cannot be expected of such an artifact would certainly be aided by a little philosophical perspective. PARTI TEN YEARS OF RESEARCH IN ARTIFICIAL INTELLIGENCE (1957-1967) 1 Phase I (1957-1962) Cognitive Simulation I. Analysis of Work in Language Translation, Problem Solving, and Pattern Recognition LANGUAGE TRANSLATION The attempts at language translation by computers had the earliest success, the most extensive and expensive research, and the most un- equivocal failure. It was soon clear that a mechanical dictionary could easily be constructed in which linguistic items, whether they were parts of words, whole words, or groups of words, could be processed indepen- dently and converted one after another into corresponding items in another language. Anthony Oettinger, the first to produce a mechanical dictionary (1954), recalls the climate of these early days: "The notion of . . . fully automatic high quality mechanical translation, planted by overzealous propagandists for automatic translation on both sides of the Iron Curtain and nurtured by the wishful thinking of potential users, blossomed like a vigorous weed." l This initial enthusiasm and the subsequent disillusionment provide a sort of paradigm for the field. It is aptly described by Bar-Hillel in his report "The Present Status of Auto- matic Translation of Languages." Notes begin on p. 221. [Citations are indicated by a superior figure. Substantive notes are indicated by a superior figure and an asterisk.] / 3 What Computers Can't Do / 4 During the first year of the research in machine translation, a considerable amount of progress was made It created among many of the workers actively engaged in this field the strong feeling that a working system was just around the corner. Though it is understandable that such an illusion should have been formed at the time, it was an illusion. It was created ... by the fact that a large number of problems were rather readily solved It was not sufficiently realized that the gap between such output . . . and high quality translation proper was still enormous, and that the problems solved until then were indeed many but just the simplest ones whereas the "few" remaining problems were the harder ones very hard indeed. 2 During the ten years following the development of a mechanical dic- tionary, five government agencies spent about $20 million on mechanical translation research. 3 In spite of journalistic claims at various moments that machine translation was at last operational, this research produced primarily a much deeper knowledge of the unsuspected complexity of syntax and semantics. As Oettinger remarks, "The major problem of selecting an appropriate target correspondent for a source word on the basis of context remains unsolved, as does the related one of establishing a unique syntactic structure for a sentence that human readers find unambiguous." 4 Oettinger concludes: "The outlook is grim for those who still cherish hopes for fully automatic high-quality mechanical transla- tion." 5 * That was in 1963. Three years later, a government report, Language and Machines, distributed by the National Academy of Sciences Na- tional Research Council, pronounced the last word on the translation boom. After carefully comparing human translations and machine products the committee concluded: We have already noted that, while we have machine-aided translation of general scientific text, we do not have useful machine translation. Furthermore, there is no immediate or predictable prospect of useful machine translation. 6 Ten years have elapsed since the early optimism concerning machine translation. At that time, flight to the moon was still science fiction, and the mechanical secretary was just around the corner. Now we have landed on the moon, and yet machine translation of typed scientific texts Phase I (1957-1962) Cognitive Simulation / 5 let alone spoken language and more general material is still over the horizon, and the horizon seems to be receding at an accelerating rate. Since much of the hope for robots like those of 2007, or for more modest servants, depends on the sort of understanding of natural language which is also necessary for machine translation, the conclusion of the National Academy of Sciences strikes at all predictions such as Minsky's that within a generation the problem of creating artificial intelligence will be substantially solved. PROBLEM SOLVING Much of the early work in the general area of artificial intelligence, especially work on game playing and problem solving, was inspired and dominated by the work of Newell, Shaw, and Simon at the RAND Corporation and at Carnegie Institute of Technology. 7 Their approach is called Cognitive Simulation (CS) because the technique generally em- ployed is to collect protocols from human subjects, which are then analyzed to discover the heuristics these subjects employ. 8 * A program is then written which incorporates similar rules of thumb. Again we find an early success: in 1957 Newell, Shaw, and Simon's Logic Theorist, using heuristically guided trial-and-error search, proved 38 out of 52 theorems from Principle, Mathematics Two years later, another Newell, Shaw, and Simon program, the General Probl im Solver (GPS), using more sophisticated means-ends analysis, solved the "canni- bal and missionary" problem and other problems of similar com- plexity. 9 * In 1961, after comparing a machine trace (see Figure 1, p. 7) with a protocol and finding that they matched to some extent, Newell and Simon concluded rather cautiously: The fragmentary evidence we have obtained to date encourages us to think that the General Problem Solver provides a rather good first approximation to an information processing theory of certain kinds of thinking and problem-solving behavior. The processes of "thinking" can no longer be regarded as completely mysterious. 10 What Computers Can't Do / 6 Soon, however, Simon gave way to more enthusiastic claims: Subsequent work has tended to confirm [our] initial hunch, and to demonstrate that heuristics, or rules of thumb, form the integral core of human problem- solving processes. As we begin to understand the nature of the heuristics that people use in thinking the mystery begins to dissolve from such (heretofore) vaguely understood processes as "intuition'* and "judgment." 11 But, as we have seen in the case of language translating, difficulties have an annoying way of reasserting themselves. This time, the "mystery" of judgment reappears in terms of the organizational aspect of the problem- solving programs. Already in 1961 at the height of Simon's enthusiasm, Minsky saw the difficulties which would attend the application of trial- and-error techniques to really complex problems: The simplest problems, e.g., playing tic-tac-toe or proving the very simplest theorems of logic, can be solved by simple recursive application of all the avail- able transformations to all the situations that occur, dealing with sub-problems in the order of their generation. This becomes impractical in more complex problems as the search space grows larger and each trial becomes more expensive in time and effort. One can no longer afford a policy of simply leaving one unsuccessful attempt to go on to another. For, each attempt on a difficult prob- lem will involve so much effort that one must be quite sure that, whatever the outcome, the effort will not be wasted entirely. One must become selective to the point that no trial is made without a compelling reason. . . , 12 This, Minsky claims, shows the need for a planning program, but as he goes on to point out: Planning methods . . . threaten to collapse when the fixed sets of categories adequate for simple problems have to be replaced by the expressions of descrip- tive language. 13 In "Some Problems of Basic Organization in Problem-Solving Pro- grams" (December 1962), Newell discusses some of the problems which arise in organizing the Chess Program, the Logic Theorist, and especially the GPS with a candor rare in the field, and admits that "most of [these problems] are unsolved to some extent, either completely, or because the solutions that have been adopted are still unsatisfactory in one way or another." 14 No further progress has been reported toward the successful Phase I (1957-1962) Cognitive Simulation / 1 LO ~<~Q-P) LI (RD~PM~RDQ) GOAL 1 TRANSFORM LI INTO LO COAL 2 DELETE R FROM LI GOAL 3 APPLY R8 TO Ll PRODUCES L2 RD-P GOAL 4 TRANSFORM L2 INTO LO GOAL 5 ADD Q TO L2 REJECT GOAL 2 GOAL 6 APPLY R8 TO Ll PRODUCES L3 -RDQ GOAL 7 TRANSFORM L3 INTO LO GOAL 8 ADD P TO L3 REJECT GOAL 2 GOAL 9 APPLY R7 TO Ll GOAL 10 CHANGE CONNECTIVE TO V IN LEFT Ll GOAL 11 APPLY R6 TO LEFT Ll PRODUCES L4 (-RV -P) (-RDQ) GOAL 12 APPLY R7 TO L4 GOAL 13 CHANGE CONNECTIVE TO V IN RIGHT L4 GOAL 14 APPLY R6 TO RIGHT L4 PRODUCES L5 (~R V ~P) (R V Q) COAL 15 APPLY R7 TO L5 GOAL 16 CHANGE SIGN OF LEFT RIGHT L5 GOAL 17 APPLY R6 TO RIGHT L5 PRODUCES L6 (~R V ~PK~R=>Q. GOAL 18 APPLY R7 TO L6 GOAL 19 CHANGE CONNECTIVE TO Y IN RIGHT L6 REJECT GOAL 16 NOTHING MORE GOAL 13 NOTHING MORE GOAL 10 NOTHING MORE Figure 1 What Computers Can't Do / 8 hierarchical organization of heuristic programs. (Significantly, the great- est achievement in the field of mechanical theorem-proving, Wang's theorem-proving program, which proved in less than five minutes all 52 theorems chosen by Newell, Shaw, and Simon, does not use heuristics.) Public admission that GPS was a dead end, however, did not come until much later. In 1967, the tenth anniversary of Simon's predictions, Newell (and Ernst) soberly, quietly, and somewhat ambiguously an- nounced that GPS was being abandoned. The preface to their paper reveals that peculiar mixture of impasse and optimism which we have begun to recognize as characteristic of the field: We have used the term "final" in several places above. This does not indicate any feeling that this document marks a terminus to our research on general problem solvers; quite the contrary is true. However, we do feel that this particu- lar aggregate of IPL-V code should be laid to rest. 15 That GPS has collapsed under the weight of its own organization becomes clearer later in the monograph. The section entitled "History of GPS" concludes: One serious limitation of the expected performance of GPS is the size of the program and the size of its rather elaborate data structure. The program itself occupies a significant portion of the computer memory and the generation of new data structures during problem solving quickly exhausts the remaining memory. Thus GPS is only designed to solve modest problems whose representation is not too elaborate. Although larger computers' memories would alleviate the extrava- gances of GPS's use of memory, conceptual difficulties would remain. 16 This curve from success to optimism to failure can be followed in miniature in the case of Gelernter's Geometry Theorem Machine (1959). Its early success with theorems like the ports asinorum gave rise to the first prediction to be totally discredited. In an article published in 1960, Gelernter explains the heuristics of his program and then concludes: "Three years ago, the dominant opinion was that the geometry machine would not exist today. And today, hardly an expert will contest the assertion that machines will be proving interesting theorems in number theory three years hence," that is, in 1963. 17 There has been no further word from Gelernter and no further progress in purely mechanical math- Phase I (1957-1962) Cognitive Simulation / 9 ematics. No more striking example exists of an "astonishing" early success and an even more astonishing failure to follow it up. PATTERN RECOGNITION This field is discussed last because the resolution of the difficulties which have arrested development in game playing, language translation, and problem solving presupposes success in the field of pattern recogni- tion (which in turn suffers from each of the difficulties encountered in the other fields). As Selfridge and Neisser point out in their classic article "Pattern Recognition by Machine," a man is continually exposed to a welter of data from his senses, and abstracts from it the patterns relevant to his activity at the moment. His ability to solve problems, prove theorems and generally run his life depends on this type of perception. We suspect that until programs to perceive patterns can be devel- oped, achievements in mechanical problem-solving will remain isolated technical triumphs. 18 There has as usual been some excellent early work. For example, the Lincoln Laboratory group under Bernard Gold produced a program for transliterating hand-sent Morse code. More recently, programs have been written for recognizing a limited set of handwritten words and printed characters in various type fonts. These all operate by searching for predetermined topological features of the characters to be recognized, and checking these features against preset or learned "definitions" of each letter in terms of these traits. The trick is to find relevant features, that is, those that remain generally invariant throughout variations of size and orientation, and other distortions. This approach has been sur- prisingly successful where recognition depends on a small number of specific traits. But none of these programs constitutes a breakthrough in pattern recognition. Each is a small engineering triumph, an ad hoc solution of a specific problem, without general applicability. As Murray Eden, who has done some of the best work in pattern recognition, summed up the situation in 1968: What Computers Can't Do / 10 Where there have been successes in performing pattern-recognition tasks by mechanical means, the successes have rested on rules that were prescribed ad hoc, in the literal sense of that phrase; that is to say, the successful methods classify reliably that particular set of patterns for which the methods were designed, but are likely to lack any significant value for classifying any other set of patterns. 19 Even in these special cases, as Selfridge and Neisser remark, "The only way the machine can get an adequate set of features is from a human programmer." 20 They thus conclude their survey of the field with a challenge rather than a prediction: The most important learning process of all is still untouched: No current pro- gram can generate test features of its own. The effectiveness of all of them is forever restricted by the ingenuity or arbitrariness of their programmers. We can barely guess how this restriction might be overcome. Until it is, 'artificial intelli- gence* will remain tainted with artifice. 21 Even these remarks may be too optimistic, however, in their supposi- tion that the present problem is feature-generation. The relative success of the Uhr-Vossler program, which generates and evaluates its own operators, shows that this problem is partially soluble. 22 But as long as recognition depends on a limited set of features, whether ad hoc or general, preprogrammed or generated, mechanical recognition has gone about as far as it can go. The number of traits that can be looked up in a reasonable amount of time is limited, and present programs have already reached this technological limit. In a paper presented at the Hawaii International Conference on the Methodologies of Pattern Rec- ognition (1968), Laveen Kanal and B. Chandrasekaran summed up the impasse as follows: Obviously, the engineering approach has built in limitations. There is a certain level of complexity above which the engineer's bag of tricks fails to produce results. As an example while even multifont printed character recognition has been successfully handled, a satisfactory solution of cursive script recognition defies all attempts. Similarly there seems to be a fairly big jump between isolated speech recognition and continuous speech recognition. Those who have been hoping to model human recognition processes have also reached an impasse. It is probable that those problems which the engineers have found difficult to handle are precisely those which have to await more detailed understanding of Phase I (1957-1962) Cognitive Simulation / 11 human recognition systems. In any case, these feelings of crisis are intimately related to those in other aspects of artificial intelligence: game playing and mechanical translation. 23 Again we find the same pattern of optimism followed by disillusion- ment. Often the disillusioned do not even understand why their hopes have been dashed, and their questioning goes unheard amidst the prom- ises and announcements of small technological advances. Such a dis- senter is Vincent Giuliano, formerly of Arthur D. Little Corporation. If Giuliano had a more detailed and insightful account of what went wrong, he would be the Oettinger or Bar-Hillel of the pattern recognition field. Like many of my colleagues, I was in hot pursuit of ways to develop something we sometimes refer to as artificial intelligence. ... in the mid-fifties, many ambitious research projects were launched with the goal of clearly demonstrating the learning capabilities of computers so that they could translate idiomatically, carry on free and natural conversations with humans, recognize speech and print it out, and diagnose diseases. All of these activities involve the discovery and learning of complex patterns. Only a few years ago we really believed that ultimately computers could be given the entire task of solving such problems, if only we could find the master key to making them do so. Alas! I feel that many of the hoped-for objectives may well be porcelain eggs; they will never hatch, no matter how long heat is applied to them, because they require pattern discovery purely on the part of machines working alone. The tasks of discovery demand human qualities. 24 Conclusion By 1962, if we are to judge by published results, an overall pattern had begun to take shape, although in some cases it was not recognized until later: an early, dramatic success based on the easy performance of simple tasks, or low-quality work on complex tasks, and then diminishing re- turns, disenchantment, and, in some cases, pessimism. This pattern is not the result of overenthusiastic pressure from eager or skeptical outsiders who demand too much too fast. The failure to produce is measured solely against the expectations of those working in the field. When the situation is grim, however, enthusiasts can always fall back What Computers Can't Do / 12 on their own optimism. This tendency to substitute long-range for opera- tional programs slips out in Feigenbaum and Feldman's claim that "the forecast for progress in research in human cognitive processes is most encouraging." 25 The forecast always has been, but one wonders: how encouraging are the prospects? Feigenbaum and Feldman claim that tangible progress is indeed being made, and they define progress very carefully as "displacement toward the ultimate goal." 26 According to this definition, the first man to climb a tree could claim tangible progress toward reaching the moon. Rather than climbing blindly, it is better to look where one is going. It is time to study in detail the specific problems confronting work in artificial intelligence and the underlying difficulties that they reveal. II. The Underlying Significance of Failure to Achieve Predicted Results Negative results, provided one recognizes them as such, can be interest- ing. Diminishing achievement, instead of the predicted accelerating suc- cess, perhaps indicates some unexpected phenomenon. Perhaps we are pushing out on a continuum like that of velocity, where further accelera- tion costs more and more energy as we approach the speed of light, or perhaps we are instead facing a discontinuity, which requires not greater effort but entirely different techniques, as in the case of the tree-climbing man who tries to reach the moon. It seems natural to take stock of the field at this point, yet surprisingly no one has done so. If someone had, he might have found that each of the four areas considered presupposes a specific form of human "infor- mation processing" that enables human subjects in that area to avoid the difficulties an artificial "subject" must confront. This section will isolate these four human forms of "information processing" and contrast them with their machine surrogates. FRINGE CONSCIOUSNESS VS. HEURISTICALLY GUIDED SEARCH It is common knowledge that certain games can be worked through on present-day computers with present-day techniques games like nim Phase I (1957-1962) Cognitive Simulation / 13 and tic-tac-toe can be programmed so that the machine will win or draw every time. Other games, however, cannot be solved in this way on present-day computers, and yet have been successfully programmed. In checkers, for example, it turns out that there are reliable ways to deter- mine the probable value of a move on the basis of certain parameters such as control of center position, advancement, and so forth. This, plus the fact that there are relatively few moves since pieces block each other and captures are forced, makes it possible to explore all plausible moves to a depth of as many as twenty moves, which proves sufficient for excellent play. Chess, however, although decidable in principle by counting out all possible moves and responses, presents the problem inevitably connected with choice mazes: exponential growth. Alternative paths multiply so rapidly that we cannot even run through all the branching possibilities far enough to form a reliable judgment as to whether a given branch is sufficiently promising to merit further exploration. Newell notes that it would take much too long to find an interesting move if the machine had to examine the possible moves of each of the pieces on the board one after another. He is also aware that if this is not done, the machine may sometimes miss an important and original combination. "We do not want the machine to spend all its time examining the future actions of committed men; yet if it were never to do this, it could overlook real opportunities." 27 NewelFs first solution was "the random element": "The machine should rarely [that is, occasionally] search for combinations which sac- rifice a Queen." 28 But this solution is unsatisfactory, as Newell himself, presumably, now realizes. The machine should not look just every once in a while for a Queen sacrifice but, rather, look in those situations in which such a sacrifice would be relevant. This is what the right heuristics are supposed to assure, by limiting the number of branches explored while retaining the more promising alternatives. But no master-level heuristics have as yet been found. All current heuristics either exclude some moves masters would find or leave open the risk of exponential growth. Simon is nonetheless convinced, for reasons to be discussed in Part II, that chess masters use such heuristics, What Computers Can't Do / 14 and so he is confident that if we listen to their protocols, follow their eye movements, perhaps question them under bright lights, we can eventu- ally discover these heuristics and build them into our program thereby pruning the exponential tree. But let us examine more closely the evi- dence that chess playing is governed by the use of heuristics. Consider the following protocol quoted by Simon, noting especially how it begins rather than how it ends. The subject says, Again I notice that one of his pieces is not defended, the Rook, and there must be ways of taking advantage of this. Suppose now, if I push the pawn up at Bishop four, if the Bishop retreats I have a Queen check and I can pick up the Rook. If, etc., etc. 29 At the end we have an example of what I shall call "counting out" thinking through the various possibilities by brute-force enumeration. We have all engaged in this process, which, guided by suitable heuristics, is supposed to account for the performance of chess masters. But how did our subject notice that the opponent's Rook was undefended? Did he examine each of his opponent's pieces and their possible defenders sequentially (or simultaneously) until he stumbled on the vulnerable Rook? That would use up too many considerations, for as Newell, Shaw, and Simon remark, "The best evidence suggests that a human player considers considerably less than 100 positions in the analysis of a move," 30 and our player must still consider many positions in evaluating the situation once the undefended Rook has been discovered. We need not appeal to introspection to discover what a player in fact does before he begins to count out; the protocol itself indicates it: the subject "zeroed in" on the promising situation ("I notice that one of his pieces is not defended"). Only after the player has zeroed in on an area does he begin to count out, to test, what he can do from there. An analysis of the MacHack program written by Richard Greenblatt will illustrate this difference between the way a human being sizes up a position and the machine's brute-force counting out. Even MacHack could not look at every alternative. The program contains a plausible move generator which limits the moves considered to the more prom- ising ones. Yet in a tough spot during a tournament, the Greenblatt Phase I (1957-1962) Cognitive Simulation / 75 program once calculated for fifteen minutes and considered 26,000 alternatives, while a human player can consider only 100, or possibly 200, moves. MacHack came up with an excellent move, which is not to say a master could not have done even better; but what is significant here is not the quality of the move, but the difference between 26,000 and 200 possibilities. This order of difference suggests that when play- ing chess, human beings are doing something different than just con- sidering alternatives, and the interesting question is: what are they doing that enables them, while considering 100 or 200 alternatives, to find more brilliant moves than the computer can find working through 26,000? The human player whose protocol we are examining is not aware of having explicitly considered or explicitly excluded from consideration any of the hundreds of possibilities that would have had to have been enumerated in order to arrive at a particular relevant area of the board by counting out. Nonetheless, the specific portion of the board which finally attracts the subject's attention depends on the overall position. To understand how this is possible, we must consider what William James has called "the fringes of consciousness": the ticking of a clock which we notice only if it stops provides a simple example of this sort of marginal awareness. Our vague awareness of the faces in a crowd when we search for a friend is another, more complex and more nearly appro- priate, case. While suggesting an alternative to the explicit awareness of counting out, neither example is entirely appropriate, however. In neither of these cases does the subject make positive use of the information resting on the fringe. The chess case is best understood in terms of Michael Polanyi's general description of the power of the fringes of consciousness to con- centrate information concerning our peripheral experience. This power resides in the area which tends to function as a background because it extends indeterminately around the central object of our attention. Seen thus from the corner of our eyes, or remembered at the back of our mind, this area compellingly affects the way we see the object on which we are focusing. We may indeed go so far as to say that we are aware of this subsidiarily noticed area mainly in the appearance of the object to which we are attending. 31 * What Computers Can't Do / 16 Once one is familiar with a house, for example, to him the front looks thicker than a fagade, because he is marginally aware of the house behind. Similarly, in chess, cues from all over the board, while remaining on the fringes of consciousness, draw attention to certain sectors by making them appear promising, dangerous, or simply worth looking into. As Newell and Simon themselves note: There are concepts in human chess playing that are much more global than those above; for example, a "developed position," "control of the center," "a won position,*' "a weak king side," "a closed position." 32 Moreover, they admit that: Sometimes de Groot's subject used very global phrases such as "... and it's a won position for White," where it is not possible to see what structure or feature of the position leads to the evaluation. " This is Newell and Simon's way of saying that they see no way of analyzing this evaluation of the overall position in terms of heuristically guided counting out. And judiciously, but without seeming to realize what this does to the plausibility of Simon's predictions, Newell and Simon go on to note: To date the work on chess programs has not shed much new light on these higher-level concepts. 34 * The attitude of Newell and Simon is typically ambiguous here. Do they think that better static evaluators that is, better heuristics for generating plausible moves could simulate zeroing in? Their continued belief in the possibility of a mechanical chess master suggests they do. However, their analysis of master play, based on the work of de Groot, should be grounds for pessimism. (As we have seen, de Groot himself says he does not have much hope for substantial improvement of heuris- tic chess programs.) Newell and Simon note that De Groot finally succeeded in separating strong from weak players by using perceptual tests involving the reproduction of chess positions after brief exposure to them (3-7 seconds). The grandmaster was able to reproduce the positions Phase I (1957-1962) Cognitive Simulation / 17 perfectly, and performance degraded appreciably with decrease in chess ability. De Groot was led to propose that perceptual abilities and organization were an important factor in very good play. 35 In the article we have already discussed, chess master Hearst casts some further light on this perceptual process and why it defies program- ming: Apparently the master perceives the setup in large units, such as pawn structure of cooperating pieces. . . . When he does make an error, it is often one of putting a piece on a very desirable square for that type of position. 36 Hearst sums up his view as follows: Because of the large number of prior associations which an experienced player has acquired, he does not visualize a chess position as a conglomeration of scattered squares and wooden pieces, but as an organized pattern (like the "Gestalt," or integrated configuration, emphasized by the Gestalt psycholo- gists.) 37 Applying these ideas to our original protocol, we can conclude that our subject's familiarity with the overall chess pattern and with the past moves of this particular game enabled him to recognize the lines offeree, the loci of strength and weakness, as well as specific positions. He sees that his opponent looks vulnerable in a certain area (just as one familiar with houses in general and with a certain house sees it as having a certain sort of back), and zeroing in on this area he discovers the unprotected Rook. This move is seen as one step in a developing pattern. There is no chess program which even tries to use the past experience of a particular game in this way. Rather, each move is taken up anew as if it were an isolated chess problem found in a book. This technique is forced upon programmers, since a program which carried along infor- mation on the past position of each piece would rapidly sink under the accumulating data. What is needed is a program which selectively carries over from the past just those features which were significant in the light of its present strategy and the strategy attributed to its opponent. But since present programs embody no. long-range strategy at all, the only alternative would be to sort through all the stored information bit by bit which would be too time consuming. Without global awareness of overall What Computers Can't Do / 18 patterns there seems to be no way of avoiding the problem of exponential growth or heuristic limitations on the possibilities which can be consid- ered. 38 * Since this global form of "information processing" in which informa- tion, rather than being explicitly considered remains on the fringes of consciousness and is implicitly taken into account, is constantly at work in organizing our experience, there is no reason to suppose that in order to discover an undefended Rook our subject must have counted out rapidly and unconsciously until he arrived at the area in which he began consciously counting out. Moreover, there are good reasons to reject this assumption, since it raises more problems than it solves. If the subject has been unconsciously counting out thousands of alter- natives with brilliant heuristics to get to the point where he focuses on that Rook, why doesn't he carry on with that unconscious process all the way to the end, until the best move just pops into his consciousness? Why, if the unconscious counting is rapid and accurate, does he resort to a cumbersome method of slowly, awkwardly, and consciously count- ing things out at the particular point where he spots the Rook? Or if, on the other hand, the unconscious counting is inadequate, what is the advantage of switching to a conscious version of the same process? This sort of teleological consideration while not a proof that uncon- scious processing is nondigital does put the burden of proof on those who claim that it is or must be. And those who make this claim have brought forward no arguments to support it. There is no evidence, behavioral or introspective, that counting out is the only kind of "infor- mation processing" involved in playing chess, that "the essential nature of the task [is] search in a space of exponentially growing possibilities." 39 On the contrary, all protocols testify that chess involves two kinds of behavior: (1) zeroing in, by means of the overall organization of the perceptual field, on an area formerly on the fringes of consciousness, and which other areas still on the fringes of consciousness make interesting; and (2) counting out explicit alternatives. This distinction clarifies the early success and the later failure of work in cognitive simulation. In all game-playing programs, early success is Phase I (1957-1962) Cognitive Simulation / 19 attained by working on those games or parts of games in which heuristi- cally guided counting out is feasible; failure occurs at the point where complexity is such that global awareness would be necessary to avoid an overwhelming exponential growth of possibilities to be counted. AMBIGUITY TOLERANCE VS, CONTEXT-FREE PRECISION Work on game playing revealed the necessity of processing "informa- tion" which is not explicitly considered or excluded, that is, information on the fringes of consciousness. Work in language translation has been halted by the need for a second nonprogrammable form of "information processing": the ability to deal with situations which are ambiguous without having to transform them by substituting a precise description. We have seen that Bar-Hillel and Oettinger, two of the most respected and best-informed workers in the field of automatic language translation, agree in their pessimistic conclusions concerning the possibility of fur- ther progress in the field. Each has realized that in order to translate a natural language, more is needed than a mechanical dictionary no matter how complete and the laws of grammar no matter how so- phisticated. The order of the words in a sentence does not provide enough information to enable a machine to determine which of several possible parsings is the appropriate one, nor do the surrounding words the written context always indicate which of several possible mean- ings is the one the author had in mind. As Oettinger says in discussing systems for producing all parsings of a sentence acceptable to a given grammar: The operation of such analyzers to date has revealed a far higher degree of legitimate syntactic ambiguity in English and in Russian than has been an- ticipated. This, and a related fuzziness of the boundary between the grammatical and the non-grammatical, raises serious questions about the possibility of effec- tive fully automatic manipulations of English or Russian for any purpose of translation or information retrieval. 40 Instead of claiming, on the basis of his early partial success with a mechanical dictionary, and later, along with Kuno and others, on syntac- What Computers Can't Do / 20 tic analyzers, that in spite of a few exceptions and difficulties, the mystery surrounding our understanding of language is beginning to dissolve, Oettinger draws attention to the "very mysterious semantic processes that enable most reasonable people to interpret most reasonable sen- tences unequivocally most of the time." 41 Here is another example of the importance of fringe consciousness. Obviously, the user of a natural language is not aware of many of the cues to which he responds in determining the intended syntax and mean- ing. On the other hand, nothing indicates that he considers each of these cues unconsciously. In fact, two considerations suggest that these cues are not the sort that could be taken up and considered by a sequential or even a parallel program. 42 * First, there is Bar-HillePs argument, which we shall later study in detail (Chapter 6), that there is an infinity of possibly relevant cues. Second, even if a manageable number of relevant cues existed, they would not help us, for in order to program a computer to use such cues to determine the meaning of an utterance, we would have to formulate syntactic and semantic criteria in terms of strict rules; whereas our use of language, while precise, is not strictly rulelike. Pascal already noted that the perceptive mind functions "tacitly, naturally, and without tech- nical rules." Wittgenstein has spelled out this insight in the case of language. We are unable clearly to circumscribe the concepts we use; not because we don't know their real definition, but because there is no real "definition" to them. To suppose that there must be would be like supposing that whenever children play with a ball they play a game according to strict rules. 43 * A natural language is used by people involved in situations in which they are pursuing certain goals. These extralinguistic goals, which need not themselves be precisely stated or statable, provide some of the cues which reduce the ambiguity of expressions as much as is necessary for the task at hand. A phrase like "stay near me" can mean anything from "press up against me" to "stand one mile away," depending upon whether it is addressed to a child in a crowd or a fellow astronaut exploring the moon. Its meaning is never unambiguous in all possible Phase I (1957-1962) Cognitive Simulation / 21 situations as if this ideal of exactitude even makes sense but the meaning can always be made sufficiently unambiguous in any particular situation so as to get the intended result. Our ability to use a global context to reduce ambiguity sufficiently without having to formalize (that is, eliminate ambiguity altogether) reveals a second fundamental form of human "information processing," which presupposes the first. Fringe consciousness takes account of cues in the context, and probably some possible parsings and meanings, all of which would have to be made explicit in the output of a machine. Our sense of the situation then allows us to exclude most of these possibilities without explicit consideration. We shall call the ability to narrow down the spectrum of possible meanings as much as the situation requires "ambiguity tolerance." Since a human being using and understanding a sentence in a natural language requires an implicit knowledge of the sentence's context- dependent use, the only way to make a computer that could understand and translate a natural language may well be, as Turing suspected, to program it to learn about the world. Bar-Hillel remarks: "I do not believe that machines whose programs do not enable them to learn, in a sophis- ticated sense of this word, will ever be able to consistently produce high-quality translations." 44 When occasionally artificial intelligence en- thusiasts admit the difficulties confronting present techniques, the appeal to learning is a favorite panacea. Seymour Papert of M.I.T., for example, has recently claimed that one cannot expect machines to perform like adults unless they are first taught, and that what is needed is a machine with the child's ability to learn. This move, however, as we shall see, only evades the problem. In the area of language learning, the only interesting and successful program is Feigenbaum's EPAM (Elementary Perceiver and Memo- rizer). EPAM simulates the learning of the association of nonsense sylla- bles, which Feigenbaum calls a simplified case of verbal learning. 45 The interesting thing about nonsense syllable learning, however, is that it is not a case of language learning at all. Learning to associate nonsense syllables is, in fact, acquiring something like a Pavlovian conditioned What Computers Can't Do / 22 reflex. The experimenter could exhibit "DAX" then "JIR," or he could flash red and then green lights; as long as two such events were associated frequently enough, one would learn to anticipate the second member of the pair. In such an experiment, the subject is assumed to be completely passive. In a sense, he isn't really learning anything, but is having some- thing done to him. Whether the subject is an idiot, a child, or an adult should ideally make no difference in the case of nonsense syllable learn- ing. Ebbinghaus, at the end of the nineteenth century, proposed this form of conditioning precisely to eliminate any use of meaningful grouping or appeal to a context of previously learned associations. It is no surprise that subject protocol and machine trace most nearly match in this area. But it is a dubious triumph: the only successful case of cognitive simulation simulates a process which does not involve com- prehension, and so is not genuinely cognitive. What is involved in learning a language is much more complicated and more mysterious than the sort of conditioned reflex involved in learning to associate nonsense syllables. To teach someone the meaning of a new word, we can sometimes point at the object which the word names. Augustine, in his Confessions, and Turing, in his article on machine intelligence, assume that this is the way we teach language to children. But Wittgenstein points out that if we simply point at a table, for exam- ple, and say "brown," a child will not know if brown is the color, the size, or the shape of the table, the kind of object, or the proper name of the object. If the child already uses language, we can say that we are pointing out the color; but if he doesn't already use language, how do we ever get off the ground? Wittgenstein suggests that the child must be engaged in a "form of life" in which he shares at least some of the goals and interests of the teacher, so that the activity at hand helps to delimit the possible reference of the words used. What, then, can be taught to a machine? This is precisely what is in question in one of the few serious objections to work in artificial intelli- gence made by one of the workers himself. A. L. Samuel, who wrote the celebrated checkers program, has argued that machines cannot be intelli- gent because they can only do what they are instructed to do. Minsky Phase I (1957-1962) Cognitive Simulation / 23 dismisses this objection with the remark that we can be surprised by the performance of our machines. 46 But Samuel certainly is aware of this, having been beaten by his own checkers program. He must mean some- thing else, presumably that the machine had to be given the program by which it could win, in a different sense than children are taught to play checkers. But if this is his defense, Samuel is already answered by Mi- chael Scriven. Scriven argues that new strategies are " 'put into* the computer by the designer ... in exactly the same metaphorical sense that we put into our children everything they come up with in their later life." 47 Still, Samuel should not let himself be bullied by the philosophers any more than by his colleagues. Data are indeed put into a machine but in an entirely different way than children are taught. We have just seen that when language is taught it is not, and, as we shall see in Chapter 6, cannot be, precisely defined. Our attempts to teach meaning must be disambiguated and assimilated in terms of a shared context. Learning as opposed to memorization and repetition requires this sort of judgment. Wittgenstein takes up this question as follows: Can someone be a man's teacher in this? Certainly. From time to time he gives him the right tip. . . . This is what learning and teaching are like here. . . . What one acquires here is not a technique; one learns correct judgements. There are also rules, but they do not form a system, and only experienced people can apply them right. Unlike calculation rules. 48 * It is this ability to grasp the point in a particular context which is true learning; since children can and must make this leap, they can and do surprise us and come up with something genuinely new. The foregoing considerations concerning the essential role of context awareness and ambiguity tolerance in the use of a natural language should suggest why, after the success of the mechanical dictionary, progress has come to a halt in the translating field. Moreover, since, as we have seen, the ability to learn a language presupposes the same complex combination of the human forms of "information processing" needed to understand a language, it is hard to see how an appeal to learning can be used to bypass the problems this area must confront. What Computers Can't Do / 24 ESSENTIAL/INESSENTIAL DISCRIMINATION VS. TRIAL-AND-ERROR SEARCH Work in problem solving also encounters two functions of thought: one, elementary and piecemeal, accounts for the early success in the field; another, more complex and requiring insight, has proved intractable to stepwise programs such as Simon's General Problem Solver. For simple problems it is possible to proceed by simply trying all possible combina- tions until one stumbles on the answer. This trial-and-error search is another example of a brute-force technique like counting out in chess. But, just as in game playing, the possibilities soon get out of hand. In problem solving one needs some systematic way to cut down the search maze so that one can spend one's time exploring promising alternatives. This is where people rely on insight and where programmers run into trouble. If a problem is set up in a simple, completely determinate way, with an end and a beginning and simple, specifically defined operations for getting from one to the other (in other words, if we have what Simon calls a "simple formal problem"), then Simon's General Problem Solver can, by trying many possibilities, bring the end and the beginning closer and closer together until the problem is solved. This would be a successful example of means-ends analysis. But even this simple case presents many difficulties. Comparing the machine print-out of the steps of a GPS solution with the transcript of the verbal report of a human being solving the same problem reveals steps in the machine trace (explicit searching) which do not appear in the subject's protocol. And Simon asks us to accept the methodologically dubious explanation of the missing steps in the human protocol that "many things concerning the task surely oc- curred without the subject's commenting on them (or being aware of them)" 49 and the even more arbitrary assumption that these further operations were of the same elementary sort as those verbalized. In fact, certain details of Newell and Simon's article, "GPS: A Program That Simulates Human Thought," suggest that these further operations are not like the programmed operations at all. Phase I (1957-1962) Cognitive Simulation 725 In one of Simon's experiments, subjects were given problems in formal logic and a list of rules for transforming symbolic expressions and asked to verbalize their attempt to solve the problems. The details of the rules are not important; what is important is that at a point in the protocol the subject notes that he applies the rule (A B ^ A) and the rule (A B VB), to the conjunction ( -R v -P) - (R v Q). Newell and Simon comment: The subject handled both forms of rule 8 together, at least as far as his comment is concerned. GPS, on the other hand, took a separate cycle of consideration for each form. Possibly the subject followed the program covertly and simply re- ported the two results together. 50 Possibly, however, the subject grasped the conjunction as symmetric with respect to the transformation operated by the rule, and so in fact applied both forms of the rule at once. Even Newell and Simon admit that they would have preferred that GPS apply both forms of the rule in the same cycle. Only then would their program provide a psychologi- cal theory of the steps the subject was going through. They wisely refrain, however, from trying to write a program which could discriminate be- tween occasions when it was appropriate to apply both forms of the rule at once and those occasions when it was not. Such a program, far from eliminating the above divergence, would require further processing not reported by the subject, thereby increasing the discrepancy between the program and the protocol. Unable thus to eliminate the divergence and unwilling to try to understand its significance, Newell and Simon dismiss the discrepancy as "an example of parallel processing." 51 * Another divergence noted by Newell and Simon, however, does not permit such an evasion. At a certain point, the protocol reads: "... I should have used rule 6 on the left-hand side of the equation. So use 6, but only on the left-hand side." Simon notes: Here we have a strong departure from the GPS trace. Both the subject and GPS found rule 6 as the appropriate one to change signs. At this point GPS simply applied the rule to the current expression; whereas the subject went back and corrected the previous application. Nothing exists in the program that corre- What Computers Can't Do / 26 spends to this. The most direct explanation is that the application of rule 6 in the inverse direction is perceived by the subject as undoing the previous applica- tion of rule 6." This is indeed the most direct explanation, but Newell and Simon do not seem to realize that this departure from the trace, which cannot be explained away by parallel processing, is as detrimental to their theory as were the discrepancies in the movements of the planets to the Ptolemaic system. Some form of thinking other than searching is taking place! Newell and Simon note the problem: "It clearly implies a mechanism (maybe a whole set of them) that is not in GPS," 53 but, like the ancient astronomers, they try to save their theory by adding a few epicycles. They continue to suppose, without any evidence, that this mechanism is just a more elaborate search technique which can be accommodated by providing GPS with "a little continuous hindsight about its past ac- tions." 54 They do not realize that their assumption that intelligent behav- ior is always the result of following heuristic rules commits them to the implausible view that their subject's decision to backtrack must be the result of a very selective checking procedure. Otherwise, all past steps would have to be rechecked at each stage, which would hopelessly en- cumber the program. A more scientific approach would be to explore further the implica- tions of the five discrepancies noted in the article, in order to determine whether or not a different form of "information processing" might be involved. For example, Gestalt pyschologist Max Wertheimer points out in his classic work, Productive Thinking, that the trial-and-error account of problem solving excludes the most important aspect of problem- solving behavior, namely a grasp of the essential structure of the prob- lem, which he calls "insight." 55 In this operation, one breaks away from the surface structure and sees the basic problem what Wertheimer calls the "deeper structure" which enables one to organize the steps neces- sary for a solution. This gestaltist conception may seem antithetical to the operational concepts demanded by artificial intelligence, but Minsky recognizes the same need in different terms: Phase I (1957-1962) Cognitive Simulation / 27 The ability to solve a difficult problem hinges on the ability to split or transform it into problems of a lower order of difficulty. To do this, without total reliance on luck, requires some understanding of the situation. One must be able to deduce, or guess, enough of the consequences of the problem statement to be able to set up simpler models of the problem situation. The models must have enough structure to make it likely that there will be a way to extend their solutions to the original problem. 56 Since insight is necessary in solving complex problems and since what Minsky demands has never been programmed, we should not be sur- prised to find that in the work of Newell and Simon this insightful restructuring of the problem is surreptitiously introduced by the pro- grammers themselves. In The Processes of Creative Thinking, Newell, Shaw, and Simon introduce "the heuristics of planning" to account for characteristics of the subject's protocol lacking in a simple means-ends analysis. We have devised a program ... to describe the way some of our subjects handle O. K. Moore's logic problems, and perhaps the easiest way to show what is involved in planning is to describe that program. On a purely pragmatic basis, the twelve operators that are admitted in this system of logic can be put in two classes, which we shall call "essential" and "inessential" operators, respectively. Essential operators are those which, when applied to an expression, make "large" changes in its appearance change "P v P" to "P," for example. Inessential operators are those which make "small" changes e.g., change "P v Q" to "Q v P." As we have said, the distinction is purely pragmatic. Of the twelve operators in this calculus, we have classified eight as essential and four as inessen- tial. . . . Next, we can take an expression and abstract from it those features that relate only to essential changes. For example, we can abstract from "P v Q" the expression (PQ), where the order of the symbols in the latter expression is regarded as irrelevant. Clearly, if inessential operations are applied to the ab- stracted expressions, the expressions will remain unchanged, while essential operations can be expected to change them. . . . We can now set up a correspondence between our original expressions and operators, on the one hand, and the abstracted expressions and essential opera- tors, on the other. Corresponding to the original problem of transforming a into 6, we can construct a new problem of transforming a' into b', where a' and b' are the expressions obtained by abstracting a and b respectively. Suppose that we solve the new problem, obtaining a sequence of expressions, a'c'd' . . . b'. What Computers Can't Do / 28 We can now transform back to the original problem space and set up the new problems of transforming a into c, c into d, and so on. Thus, the solution of the problem in the planning space provides a plan for the solution of the original problem. 57 No comment is necessary. One merely has to note that the actual pro- gram description begins in the second paragraph. The classification of the operators into essential and inessential, the function Wertheimer calls "finding the deeper structure" or "insight," is introduced by the programmers before the actual programming begins. This sleight of hand is overlooked by Miller, Galanter, and Pribram in Plans and the Structure of Behavior, a book which presents a psycho- logical theory influenced by Newell, Shaw, and Simon's work. Miller et al. begin by quoting Polya, who is fully aware of the necessary role insight plays in problem solving: In his popular text, How to Solve It, Polya distinguishes . . . phases in the heuristic process: First, we must understand the problem. We have to see clearly what the data are, what conditions are imposed, and what the unknown thing is that we are searching for. Second, we must devise a plan that will guide the solution and connect the data to the unknown. 58 Miller et al. then minimize the importance of phase I, or rather simply decide not to worry about it. Obviously, the second of these is most critical. The first is what we have described in Chapter 12 as the construction of a clear Image of the situation in order to establish a test for the solution of the problem; it is indispensable, of course, but in the discussion of well-defined problems we assume that it has already been accomplished. 59 Still the whole psychological theory of problem solving will not be worth much if there is no way to bring step one into the computer model. Therefore, it is no surprise that ten pages later, after adopting Simon's means-ends analysis, we find Miller et al. referring with relief to Simon's "planning method," 60 presumably the very paragraphs we have just dis- cussed: Phase I (1957-1 962) Cognitive Simulation / 29 A second very general system of heuristics used by Newell, Shaw, and Simon consists in omitting certain details of the problem. This usually simplifies the task and the simplified problem may be solved by some familiar plan. The plan used to solve the simple problem is then used as the strategy for solving the original, complicated problem. In solving a problem in the propositional calculus, for example, the machine can decide to ignore differences among the logical connec- tives and the order of the symbols. . . . 6l But, as we have seen, it is not the machine that decides, but Newell, Shaw, and Simon, themselves. To speak of heuristics here is completely misleading, since no one has succeeded in formulating the rules which guide this preliminary choice or even in showing that at this stage, where insight is required, people follow rules. Thus we are left with no com- puter theory of the fundamental first step in all problem solving: the making of the essential/inessential distinction. Only those with faith such as that of Miller et al. could have missed the fact that Simon's "planning method," with its predigesting of the material, poses the prob- lem for computer simulation rather than provides the solution. This human ability to distinguish the essential from the inessential in a specific task accounts for the divergence of the protocol of the problem- solving subjects from the GPS trace. We have already suggested that the subject applies both forms of rule 8 together because he realizes at this initial stage that both sides of the conjunction are functionally equiva- lent. Likewise, because he has grasped the essential function of rule 6, the subject can see that a second application of the rule simply neutral- izes the previous one. As Wertheimer notes: The process [of structuring a problem] does not involve merely the given parts and their transformations. It works in conjunction with material that is structur- ally relevant but is selected from past experience. . . ." Since game playing is a form of problem solving we should expect to find the same process in chess playing, and indeed we do. To quote Hearst: De Groot concluded from his study that differences in playing strength depend much less on calculating power than on "skill in problem conception." Grand- masters seem to be superior to masters in isolating the most significant features of a position, rather than in the total number of moves that they consider. What Computers Can't Do / 30 Somewhat surprisingly, de Groot found that grandmasters do not examine more possibilities on a single move than lower-ranked experts or masters (an average of two to four first moves per position) nor do they look further ahead (usually a maximum of six to seven moves ahead for each). The grandmaster is somehow able to "see** the core of the problem immediately, whereas the expert or lesser player finds it with difficulty, or misses it completely, even though he analyzes as many alternatives and looks as many moves ahead as the grandmaster. 63 In 1961, as we have seen, Minsky was already aware of these problems. But his only hope was that one would discover a planning program which would use more of the same sort of heuristic searching on a higher level: When we call for the use of "reasoning," we intend no suggestion of giving up the game by invoking an intelligent subroutine. The program that administers the search will be just another heuristic program. Almost certainly it will be composed largely of the same sorts of objects and processes that will comprise the subject-domain programs. 64 But such a planning program itself would require a distinction between essential and inessential operators. Unless at some stage the programmer himself introduces this distinction, he will be forced into an infinite regress of planning programs, each one of which will require a higher- order program to structure its ill-structured data. At this point, the transition from the easy to the difficult form of "information proces- sing," Minsky makes the typical move to learning. The problem of making useful deductions from a large body of statements (e.g. about the relevance of different methods to different kinds of problems) raises a new search problem. One must restrict the logical exploration to data likely to be relevant to the current problem. This selection function could hardly be completely built in at the start. It must develop along with other data ac- cumulated by experience. 65 But thus far no one has even tried to suggest how a machine could perform this selection operation, or how it could be programmed to learn to perform it, since it is one of the conditions for learning from past experience. Feigenbaum, in a recent appraisal of work done since Computers and Thought, notes the glaring lack of learning programs: Phase I (1957-1962) Cognitive Simulation / 31 The AI field still has little grasp of the machine learning problem for problem solvers. For many years, almost the only citation worth making was to Samuel's famed checker playing program and its learning system. (Great interest arose once in a scheme proposed by Newell, Shaw, and Simon for learning in GPS, but the scheme was never realized.) Surprisingly, today we face the same situa- tion. 66 This lack of progress is surprising only to those, like Feigenbaum, who do not recognize the ability to distinguish the essential from the inessen- tial as a human form of "information processing," necessary for learning and problem solving, yet not amenable to the mechanical search tech- niques which may operate once this distinction has been made. It is precisely this function of intelligence which resists further progress in the problem-solving field. It is an illusion, moreover, to think that the planning problem can be solved in isolation; that essential/inessential operations are given like blocks and one need only sort them out. It is easy to be hypnotized by oversimplified and ad hoc cases like the logic problem into thinking that some operations are essential or inessential in themselves. It then looks as if we can find them because they are already there, so that we simply have to discover a heuristic rule to sort them out. But normally (and often even in logic) essential operations are not around to be found because they do not exist independently of the pragmatic context. In the light of their frank inevitable recourse to the insightful predi- gesting of their material, there seems to be no foundation for Newell, Shaw, and Simon's claim that the behavior vaguely labeled cleverness or keen insight in human problem solving is really just the result of the judicious application of certain heuristics for narrowing the search for solutions. Their work with GPS, on the contrary, demonstrates that all searching, unless directed by a preliminary structuring of the problem, is merely muddling through. Ironically, research in Cognitive Simulation is a perfect example of so-called intelligent behavior which proceeds like the unaided GPS. Here one finds the kind of tinkering and ad hoc patchwork characteristic of a fascination with the surface structure a sort of tree-climbing with one's eyes on the moon. Perhaps it is just because the field provides no What Computers Can't Do / 32 example of insight that some people in Cognitive Simulation have mis- taken the operation of GPS for intelligent behavior. PERSPICUOUS GROUPING VS. CHARACTER LISTS A computer must .recognize all patterns in terms of a list of specific traits. This raises problems of exponential growth which human beings are able to avoid by proceeding in a different way. Simulating recognition of even simple patterns may thus require recourse to each of the funda- mental forms of human "information processing" discussed this far. And even if in these simple cases artificial intelligence workers have been able to make some headway with mechanical techniques, patterns as complex as artistic styles and the human face reveal a loose sort of resemblance which seems to require a special combination of insight, fringe conscious- ness, and ambiguity tolerance beyond the reach of digital machines. It is no wonder, then, that work in pattern recognition has had a late start and an early stagnation. In Chapter 1 we noted that a weakness of current pattern recognition programs (with the possible exception of the Uhr-Vossler program, the power of whose operators since it only, recognizes five letters has not yet been sufficiently tested) is that they are not able to determine their own selection operators. Now, however, we shall see that this way of presenting the problem is based on assumptions that hide deeper and more difficult issues. Insight A first indication that human pattern recognition differs radi- cally from mechanical recognition is seen in human (and animal) toler- ance for changes in orientation and size, degrees of incompleteness and distortion, and amount of background noise. An early artificial intelligence approach was to try to normalize the pattern and then to test it against a set of templates to see which it matched. Human recognition, on the other hand, seems to simply disre- gard changes in size and orientation, as well as breaks in the figure, and so on. Although certain perceptual constants do achieve some normali- zation (apparent size and brightness do not vary as much as correspond- ing changes in the signal reaching the retina), it seems clear that we do Phase I (1957-1962) Cognitive Simulation / 33 not need to fully normalize and smooth out the pattern, since we can perceive the pattern as skewed, incomplete, large or small, and so on, at the same time we recognize it. More recent programs, rather than normalizing the pattern, seek pow- erful operators which pick out discriminating traits but remain insensi- tive to distortion and noise. But human beings, when recognizing patterns, do not seem to employ these artificial expedients either. In those special cases where human beings can articulate their cues, these turn out not to be powerful operators which include sloppy patterns and exclude noise, but rather a set of ideal traits which are only approximated in the specific instances of patterns recognized. Distorted patterns are recognized not as falling under some looser and more ingenious set of traits, but as exhibiting the same simple traits as the undistorted figures, along with certain accidental additions or omissions. Similarly, noise is not tested and excluded; it is ignored as inessential. 67 * Here again, we note the human ability to distinguish the essential from the inessential. Fringe Consciousness. To determine which of a set of already analyzed patterns a presented pattern most nearly resembles, workers have pro- posed analyzing the presented pattern for a set of traits by means of a decision tree; or by combining the probabilities that each of a set of traits is present, as in Selfridge's Pandaemonium program. Either method uncritically assumes that a human being, like a mechanical pattern recognizer, must classify a pattern in terms of a specific list of traits. It seems self-evident to Selfridge and Neisser that "a man who abstracts a pattern from a complex of stimuli has essentially classified the possible inputs." 68 Earl Hunt makes the same assumption in his review of pattern recognition work: "Pattern recognition, like concept learning, involves the learning of a classification rule." 69 Yet, if the pattern is at all complicated and sufficiently similar to many other patterns so that many traits are needed for discrimination, the problem of exponential growth threatens. Supposing that a trait-by-trait analysis is the way any pattern recognizer, human or artificial, must proceed thus leads to the assumption that there must be certain crucial traits if one could only find them, or program the machine to find them for itself which would make the processing manageable. What Computers Can't Do / 34 One is led to look for a sort of perceptual heuristic, the "powerful operators" which no one as yet has been able to find. And just as the chess masters are not able to provide the programmer with the heuristic shortcuts they are supposed to be using, Selfridge and Neisser note in the case of pattern recognition that "very often the basis of classification is unknown, even to [the analyzer]." Nevertheless, Selfridge and Neisser assume, like Newell and Simon, that unconsciously a maze is being explored in this case, that a list of traits is being searched. They are thus led to conclude that "it [the basis of classification] is too complex to be specified explicitly." 70 But the difficulties involved in searching such a list suggest again that, for human beings at least, not all possibly relevant traits are taken up in a series or in parallel and used to make some sort of decision, but that many traits crucial to discrimination are never taken up explicitly at all but do their work while remaining on the fringe of consciousness. Whereas in chess we begin with a global sense of the situation and have recourse to counting out only in the last analysis, in perception we need never appeal to any explicit traits. We normally recognize an object as similar to other objects without being aware of it as an example of a type or as a member of a class defined in terms of specific traits. As Aron Gurwitsch puts it in his analysis of the difference between perceptual and conceptual consciousness: Perceived objects appear to us with generic determinations. . . . But and this is the decisive point to perceive an object of a certain kind is not at all the same thing as grasping that object as representative or as a particular case of a type. 71 Of course, we can sometimes make the defining traits explicit: The first step in the constituting of conceptual consciousness consists in effecting a dissociation within the object perceived in its typicality. The generic traits which until then were immanent and inherent in the perceived thing are detached and disengaged from it. Rendered explicit, these traits can be seized in them- selves. . . . Consequent upon this dissociation, the generic becomes the general. From this aspect it opposes itself to the thing perceived from which it has just been disengaged, and which now is transformed into an example, a particular instance. . . . [Thus, cues] can be grasped and become themes [specific traits we are aware Phase I (1957-1962) Cognitive Simulation 735 of] . . . , whereas previously they only contributed to the constitution of another theme [the pattern] within which they played only a mute role. 72 This shift from perceptual to conceptual consciousness (from the per- ceptive to the mathematical frame of mind, to use Pascal's expression), is not necessarily an improvement. Certain victims of aphasia, studied by Gelb and Goldstein, have lost their capacity for perceptual recogni- tion. All recognition for the patient becomes a question of classification. The patient has to resort to checklists and search procedures, like a digital computer. Some such aphasics can only recognize a figure such as a triangle by listing its traits, that is, by counting its sides and then thinking: "A triangle has three sides. Therefore, this is a triangle." 73 Such conceptual recognition is time consuming and unwieldy; the victims of such brain injuries are utterly incapable of getting along in the everyday world. Evidently, in pattern recognition, passing from implicit perceptual grouping to explicit conceptual classification even at some final stage, as in chess is usually disadvantageous. The fact that we need not con- ceptualize or thematize the traits common to several instances of the same pattern in order to recognize that pattern distinguishes human recognition from machine recognition, which only occurs on the explicit conceptual level of class membership. Context-Dependent Ambiguity Reduction. In the cases thus far consid- ered, the traits defining a member of a class, while generally too numer- ous to be useful in practical recognition, could at least in principle always be made explicit. In some cases, however, such explicitation is not even possible. To appreciate this point we must first get over the idea, shared by traditional philosophers and workers in artificial intelligence alike, that pattern recognition can always be understood as a sort of classifica- tion. In this overhasty generalization three distinct kinds of pattern recognition are lumped together, none of which has the characteristics philosophers and digital computers demand. First there is the recognition of what Gurwitsch calls the generic. An example of such recognition would be the recognition of a certain object as a pencil. As Gurwitsch has pointed out, this form of recognition, while What Computers Can't Do / 36 not explicit, lends itself to explicitation in terms of a set of features. It might thus seem adapted to being programmed. But what Gurwitsch overlooks in his account is that in this form of recognition our purposes serve to select which features are significant, and, among these, certain features which are crucial. For example, it is significant for our purposes that a pen have a point. However, when a writing instrument with a ball at the end was introduced, the end was nonetheless called a point (not a tip), and the instrument a ball-point pen (not a pencil), presumably because it was crucial to the users that the mark this instrument made could not be erased. We might conclude that making an indelible mark is a defining crite- rion for being a pen, whereas having a point is only what Wittgenstein calls a symptom ". . . a phenomenon of which the experience has taught us that it coincided, in some way or other, with the phenomenon which is our defining criterion." We might even try to introduce this distinction between symptom and criterion into our program. But Wittgenstein's essential point in distinguishing between symptom and criterion is that the distinction is not fixed once and for all but changes with our changing purposes and knowledge: In practice, if you were asked which phenomenon is the defining criterion and which is a symptom, you would in most cases be unable to answer this question except by making an arbitrary decision ad hoc. It may be practical to define a word by taking one phenomenon as the defining criterion, but we shall easily be persuaded to define the word by means of what, according to our first use, was a symptom. Doctors will use names of diseases without ever deciding which phenomena are to be taken as criteria and which as symptoms; and this need not be a deplorable lack of clarity. 74 Indeed, it is one way our concepts gain the openness crucial to human pattern recognition, a flexibility lacking in a computer using a fixed set of essential features. A second sort of pattern recognition is the recognition of resemblance. In this sort of recognition, as in "narrowing down" 75 * the meaning of words or sentences, the context plays a determining role. The context may simply lead us to notice those resemblances which we can subse- quently recognize in isolation as in the case of ambiguous figures such Phase I (1957-1962) Cognitive Simulation / 37 as Wittgenstein's duck-rabbit, which resembles a duck when surrounded by pictures of ducks and a rabbit when surrounded by rabbits or it may lead us to focus on certain aspects of the pattern, as in Pudovkin's famous experiment: One day Pudovkin took a close-up of Mosjoukin with a completely impassive expression and projected it after showing: first, a bowl of soup, then, a young woman lying dead in her coffin, and last, a child playing with a teddy-bear. The first thing noticed was that Mosjoukin seemed to be looking at the bowl, the young woman, and the child, and next one noted that he was looking pensively at the dish, that he wore an expression of sorrow when looking at the woman, and that he had a glowing smile for the child. The audience was amazed at his variety of expression, although the same shot had actually been used all three times and was, if anything, remarkably inexpressive. 76 Here, in a striking way, the meaning of the context determines what expression is seen on the face in a situation in which no traits of the face as projected on the screen could account for these differences. Still one might say that the expressive face, the one that the viewers thought they saw, had certain traits, like sad eyes, or a happy smile, which led the viewer to recognize the expression. But the expression of a person's eyes, for example, may depend on the whole face in such a way as to be unrecognizable if viewed through a slit. Moreover, a certain expression of the eyes may bring out a certain curve of the nose which would not be noticed if the nose were in another face; the nose in turn may give a certain twist to the smile which may affect the appearance of the eyes. As Wittgenstein remarks: "A human mouth smiles only in a human face." 77 In such cases, the traits necessary for recognizing a resemblance (dancing eyes, mocking smile, etc.) cannot, even when thematized, be isolated and defined in a neutral, context-free way. Moreover, as in the case of linguistic disambiguation, the context in this example the whole face not only determines the features essential for recognition, but is reciprocally determined by them. The expression is not deduced from the traits; it is simply the organization of the eyes, the mouth, and so forth, just as a melody is made up of the very notes to which it gives their particular values. In this sort of resemblance, the notion of recognizing the pattern in terms of isolated traits makes no sense. What Computers Can't Do / 38 In another case of resemblance, objects recognized as belonging to- gether need not have any traits in common at all not even context- dependent ones. Wittgenstein, in his study of natural language, was led to investigate this type of nonclassifactory recognition: We see a complicated network of similarities overlapping and criss-crossing: Sometimes overall similarities, sometimes similarities of detail. I can think of no better expression to characterize these similarities than "family resemblances"; for the various resemblances between members of a family: build, features, color of eyes, gait, temperament, etc. etc., overlap and criss-cross in the same way. . . . We extend our concept . . as in spinning a thread we twist fiber on fiber. 78 Family resemblance differs from class membership in several impor- tant ways: classes can be defined in terms of traits even if they have no members, whereas family resemblances are recognized only in terms of real or imaginary examples. 79 * Moreover, whereas class membership is all or nothing, 80 * family resemblance allows a spectrum ranging from the typical to the atypical. An atypical member of a family, for example, may be recognized by being placed in a series of faces leading from a typical member to the atypical one. Similarly, certain concepts like graceful, garish, and crude can not be defined in terms of necessary and sufficient conditions, but only by exhibiting a typical case. Since this sort of recog- nition of a member of a "family" is accomplished not by a list of traits, but by seeing the case in question in terms of its proximity to a paradigm (i.e., typical) case, such recognition gives us another kind of openness and flexibility. Finally Wittgenstein goes even further and suggests that in some kinds of recognition there may be no common traits, even overlapping ones. Wittgenstein continues the above remarks rather obscurely: . . . If someone wishes to say: "There is something common to all these construc- tions namely the disjunction of all their common properties" I should reply: Now you are only playing with words. One might as well say: "Something runs through the whole thread namely the continuous overlapping of these fibres." 81 Wittgenstein here may be suggesting a third kind of recognition which he does not clearly distinguish from resemblance, but which we might call the recognition of similarity. Phase I (1957-1962) Cognitive Simulation / 39 Wittgenstein, on this interpretation, should not be taken to mean that recognition involves so many overlapping traits, but that one cannot use such an unwieldy disjunction. A more consistent way of understanding his analysis would be to conclude that each of the traits he mentions in discussing family resemblance the build, color of eyes, gait, etc. is not identical in any two members of the family, but in turn consists of a network of crisscrossing similarities. To follow the analogy, each fiber is made of fibers all the way down. Thus, no two members of a family need have any identical features for them all to share a family resemblance. Similarity is the ultimate notion in Wittgenstein's analysis and it cannot be reduced as machine-thinking would require to a list or disjunction of identical, determinate features. 82 Those capable of recognizing a member of a "family" need not be able to list any exactly similar traits common to even two members, nor is there any reason to suppose such traits exist. Indeed, formalizing family resemblance in terms of exactly similar traits would eliminate a kind of openness to new cases which is the most striking feature of this form of recognition. No matter what disjunctive list of traits is constructed, one can always invent a new "family" member whose traits are similar to those of the given members without being exactly similar to any of the traits of any of them, and which in some situation would be recognized as belonging with the others. This sophisticated but nonetheless very common form of recognition employs a special combination of the three forms of "information pro- cessing" discussed thus far: fringe consciousness, insight, and context dependence. To begin with, the process is implicit. It uses information which, in a manner of speaking, remains on the fringes of consciousness. To see the role of insight we must first distinguish the generic from the typical, although Gurwitsch uses these two terms interchangeably. As Gurwitsch defines it, recognition of the generic depends on implicit cues which can always be made explicit. Recognition of the typical, as we have been using the term, depends on similarities which cannot be thematized. Recognition of the typical, then, unlike recognition of the generic, re- quires insightful ordering around a paradigm. A paradigm case serves its function insofar as it is the clearest manifestation of what (essentially) What Computers Can't Do / 40 makes all members, members of a given group. Finally, recognition in terms of proximity to the paradigm is a form of context dependence. Wittgenstein remarks that "a perspicuous representation produces just that understanding which consists in seeing connections." 83 Following Wittgenstein, we have called this combination of fringe consciousness, insight, and context determination "perspicuous grouping." This form of human "information processing" is as important as the three funda- mental forms of information processing from which it is derived. Summary. Human beings are able to recognize patterns under the following increasingly difficult conditions: 1. The pattern may be skewed, incomplete, deformed, and embedded in noise; 2. The traits required for recognition may be "so fine and so numerous" that, even if they could be formalized, a search through a branching list of such traits would soon become unmanageable as new patterns for discrimina- tion were added; 3. The traits may depend upon external and internal context and are thus not amenable to context-free specification; 4. There may be no common traits but a "complicated network of overlap- ping similarities," capable of assimilating ever new variations. Any system which can equal human performance, must therefore, be able to 1. Distinguish the essential from the inessential features of a particular in- stance of a pattern; 2. Use cues which remain on the fringes of consciousness; 3. Take account of the context; 4. Perceive the individual as typical, i.e., situate the individual with respect to a paradigm case. Since the recognition of patterns of even moderate complexity may require these four forms of human "information processing," work in pattern recognition has not progressed beyond the laborious recognition of simple alphanumeric patterns such as typewriter fonts and zip code figures. Moreover, it is generally acknowledged that further progress in game playing, language translation, and problem solving awaits a break- through in pattern recognition research. Phase I (1957-1962) Cognitive Simulation / 41 Conclusion The basic problem facing workers attempting to use computers in the simulation of human intelligent behavior should now be clear: all alter- natives must be made explicit. In game playing, the exponential growth of the tree of these alternative paths requires a restriction on the paths which can be followed out; in complicated games such as chess, pro- grams cannot now select the most promising paths. In problem solving, the issue is not only how to direct a selective search among the explicit alternatives, but how to structure the problem so as to begin the search process. In language translation, even the elements to be manipulated are not clear due to the intrinsic ambiguities of a natural language; in pattern recognition, all three difficulties are inextricably intertwined, as well as the fact that similarity and typicality seem to be irreducible characteris- tics of perception. These difficulties have brought to a standstill the first five years of work on Cognitive Simulation. None of Simon's predictions has been fulfilled. The failure to fulfill the first two, about how well machines could do in chess and mathematics, gave the lie to Simon's third prediction concerning a psychological the- ory of human behavior. In spite of the eagerness and gullibility of psy- chologists, within the past ten years most theories in psychology have not taken the form of computer programs. Instead of these triumphs, an overall pattern has emerged: success with simple mechanical forms of information processing, great expecta- tions, and then failure when confronted with more complicated forms of behavior. Simon's predictions fall into place as just another example of the phenomenon which Bar-Hillel has called the "fallacy of the success- ful first step." 84 * Simon himself, however, has drawn no such sobering conclusions. In his latest prediction, made in 1965, Simon now affirms that "machines will be capable, within twenty years, of doing any work that a man can do." 85 We shall devote Part II to the reasons for this imperturbable optimism, but first we must consider the work in AI which has taken up where work in Cognitive Simulation gave out. Phase II (1962-1967) Semantic Information Processing To place Phase I in perspective and to form an idea of what was expected and accomplished in Phase II, it is helpful to begin by quot- ing Minsky's brief account of the history of work on machine intelli- gence: In the early 1950's, as general-purpose computers became available to the scien- tific community, Cybernetics divided . . . into three chief avenues: The first was the continuation of the search for simple basic principles. This became trans- formed into the goal of discovering what we might call minimal, Self-Organizing Systems. A paradigm of this approach is to find large collections of generally similar components that, when arranged in a very weakly specified structure and placed in an appropriate environment, would eventually come to behave in an "adaptive" fashion. Eventually, it was hoped, intelligent behavior would emerge from the evolution of such a system. 1 Since those still pursuing this course, sometimes called cybernetics, have produced no interesting results although their spokesman, Frank Rosenblatt, has produced some of the most fantastic promises and claims 2 * they will not be dealt with here. The second important avenue was an attempt to build working models of hu- man behavior, . . . requiring the machine's behavior to match that of human sub- jects 3 Phase II (1962-1967) Semantic Information Processing / 43 The book, Computers and Thought, edited by E. Feigenbaum and J. Feldman who did their graduate work in the Carnegie group, gives a good view of the state of affairs as it stood by about the end of 1961. 4 This is the research in Cognitive Simulation, led by Newell and Simon, which we have criticized in Chapter 1. Minsky is similarly critical of this work in a paper delivered at the time Phase I was nearing its end: Methods that worked quite well on easy problems did not extend smoothly to the difficult ones. Continued progress will require implementation of new ideas, for there are some very tough problems in our immediate path. 5 This is Minsky's way of recognizing the stagnation we have noted. At the same time Minsky and his group at M.I.T. undertook to provide new ideas and their implementation: The third approach, the one we call Artificial Intelligence, was an attempt to build intelligent machines without any prejudice toward making the system simple, biological, or humanoid. Workers taking this route regarded the self- organizing systems as unpromising or, perhaps, premature. Even if simplicity of initial organization was to be an ultimate goal, one might first need experience with working intelligent systems (based if necessary on ad hoc mechanisms) if one were eventually to be able to design more economical schemes. 6 We shall now turn to this third and most recent approach, the results of which are reported in Minsky's book Semantic Information Process- ing, to see just what has actually been accomplished. Minsky once sug- gested that in evaluating the programs presented in his book one might ask five questions: 1. Why were these particular problems selected? 2. How do these programs work? 3. What are their limitations? 4. What do the programs actually achieve? 5. How can they be extended to larger domains of competence? If, following this method, we analyze the programs which Minsky pre- sents as the best work since 1962, we shall find that unlike work done before 1961, which tended to give the impression of intelligence by simulating simple, mechanical aspects of intelligent behavior, the current approach is characterized by ad hoc solutions of cleverly chosen problems, What Computers Can't Do / 44 which give the illusion of complex intellectual activity. In fact, however, problems which arrested work in 1961 still remain unsolved. We shall also find again that only an unquestioned underlying faith enables work- ers such as Minsky to find this situation encouraging. Let us look at the programs in detail. I. Analysis of Semantic Information Processing Programs ANALYSIS OF A PROGRAM WHICH "UNDERSTANDS ENGLISH" BOBROW'S STUDENT Of the five semantic information processing programs collected in Minsky's book, Daniel Bobrow's STUDENT a program for solving algebra word problems is put forward as the most successful. It is, Minsky tells us, "a demonstration par excellence of the power of using meaning to solve linguistic problems." 7 Indeed, Minsky devotes a great deal of his Scientific American article to Bobrow's program and goes so far as to say that "it understands English." 8 Since this program is presented as the best so far, we shall begin by analyzing it in detail, according to Minsky's suggested five ques- tions. First: Why was this particular problem selected? Bobrow himself tells us: In constructing a question-answering system many problems are greatly sim- plified if the problem context is restricted. 9 Moreover, There are a number of reasons for choosing the context of algebra story problems in which to develop techniques which would allow a computer problem solving system to accept a natural language input. First, we know a good type of data structure in which to store information needed to answer questions in this context, namely, algebraic equations. 10 It is important to note that the problem was chosen because the restricted context made it easier. The full significance of this restriction, Phase I! (1962-1967) Semantic Information Processing / 45 however, will only be evident after we have answered the next two questions. How does the program work? The program simply breaks up the sentences of the story problem into units on the basis of cues such as the words "times," "of," "equals," etc.; equates these sentence chunks with x's and /s; and tries to set up simultaneous equations. If these equations cannot be solved, it appeals to further rules for breaking up the sentences into other units and tries again. The whole scheme works only because there is the constraint, not present in understanding ordinary discourse, that the pieces of the sen- tence, when represented by variables, will set up soluble equations. As Minsky puts it: ". . . some possibly syntactic ambiguities in the input are decided on the overall basis of algebraic consistency. . . ." H Choosing algebra problems also has another advantage: In natural language, the ambiguities arise not only from the variety of structural groupings the words could be given, but also from the variety of meanings that can be assigned to each individual word. In STUDENT the strong semantic constraint (that the sentences express algebraic relations between the designated entities) keeps the situation more or less under control. 12 What are the limitations of the program? The advantage of using algebraic constraints is also a serious limita- tion on the generality of the program, however, for such a "strong constraint" eliminates just that aspect of natural language, namely its ambiguity, which makes machine processing of natural language diffi- cult, if not impossible. Such a program is so far from semantic under- standing that, as Bobrow admits, ". . . the phrase 'the number of times I went to the movies' which should be interpreted as a variable string will be interpreted incorrectly as the product of the two variables 'number of and 'I went to the movies,' because 'times' is always considered to be an operator." 13 What, then, has been achieved? Bobrow is rather cautious. Although his thesis is somewhat mislead- ingly entitled "Natural Language Input for a Computer Problem Solving What Computers Can't Do / 46 Program," Bobrow makes clear from the outset that the program "ac- cepts as input a comfortable but restricted subset of English. " M He adds: In the following discussion, I shall use phrases such as "the computer under- stands English." In all such cases, the "English" is just the restricted subset of English allowable as input for the computer program under discussion. 15 This is straightforward enough, and seems an admirable attempt to claim no more than is justified by the restricted choice of material. In the course of the work, Bobrow even makes clear that "The STUDENT program considers words as symbols, and makes do with as little knowl- edge about the meaning of words as is compatible with the goal of finding a solution of the particular problem." 16 In other words this program embodies a minimum of semantic under- standing. Bobrow is proud that he can get so much for so little: "The semantic model in the STUDENT system is based on one relationship (equality) and five basic arithmetic functions." 17 Bobrow is equally careful in noting he has given a special meaning to "understands." For purposes of this report I have adopted the following operational definition of "understanding." A computer understands a subset of English if it accepts input sentences which are members of this subset, and answers questions based on information contained in the input. The STUDENT system understands English in this sense. 18 * Bobrow concludes cautiously: "I think we are far from writing a pro- gram which can understand all, or even a very large segment, of English. However, within its narrow field of competence, STUDENT has demon- strated that 'understanding' machines can be built." 19 Yet Minsky says in his Scientific American article that "STUDENT . . . understands English." What has happened? Bobrow's quotation marks around "understanding" are the key. If we remember that "understands" merely means "answers questions in a restricted subset of English subject to algebraic constraints," then we will also remember that although the words in quotation marks have nothing Phase II (1962-1967) Semantic Information Processing / 47 to do with what human understanding normally means, they are nonetheless accurate. However, one can't help being misled into feeling that if Bobrow uses "understands" rather than "processes," it must be because his program has something to do with human understanding. Minsky exploits this ambiguity in his rhetorical article simply by drop- ping the quotation marks. Minsky makes even more surprising and misleading claims concerning the "enormous 'learning potential' " of Bobrow's program: Consider the qualitative effect, upon the subsequent performance of Bobrow's STUDENT, of telling it that "distance equals speed times timer That one experience alone enables it to handle a large new portion of "high-school alge- bra"; the physical position-velocity-time problems. It is important not to fall into the habit ... of concentrating only on the kind of "learning" that appears as slow-improvement-attendant-upon-sickeningly-often-repeated experience! Bobrow's program does not have any cautious statistical devices that have to be told something over and over again, so its learning is too brilliant to be called so. 20 Again it is easy to show that what has been acquired by the machine can in no way be called "understanding." The machine has indeed been given another equation, but it does not understand it as a formula. That is, the program can now plug one distance, one rate, and one time into the equation d = rt; but that it does not understand anything is clear from the fact that it cannot use this equation twice in one problem, for it has no way of determining which quantities should be used in which equation. As Bobrow admits: "the same phrase must always be used to represent the same variable in a problem." 21 No learning has occurred. Once he has removed the quotation marks from "understand" and interpreted the quotation marks around "learning" to mean superhuman learning, Minsky is free to engage in the usual riot of speculation. In order for a program to improve itself substantially it would have to have at least a rudimentary understanding of its own problem-solving process and some ability to recognize an improvement when it found one. There is no inherent reason why this should be impossible for a machine. Given a model of its own What Computers Can't Do / 48 workings, it could use its problem-solving power to work on the problem of self-improvement. . . . Once we have devised programs with a genuine capacity for self-improvement a rapid evolutionary process will begin. As the machine improves both itself and its model of itself, we shall begin to see all the phenomena associated with the terms "consciousness," "intuition" and "intelligence" itself. It is hard to say how close we are to this threshold, but once it is crossed the world will not be the same. 22 It is not as hard to say how close we are to this threshold as Minsky would like us to believe. Since the success of Bobrow's program has allegedly given us the rudiments of understanding and learning that Minsky is relying on, we need only ask: to what extent can Bobrow's techniques be generalized and extended? Which leads us to question five: How can the program in question be extended to larger domains of competence? Here even Bobrow throws his caution to the winds and in spite of his earlier remark that the semantic model is based on one relationship (equality); that is, only sets up and solves equations where it can use the algebraic constraint claims that his "semantic theory of discourse can be used as a basis for a much more general language processing sys- tem." 23 And Bobrow concludes the abstract of his thesis with the now familiar first-step fallacy: "The STUDENT system is a first step toward natural language communication with computers. Further work on the semantic theory proposed should result in much more sophisticated systems." 24 Five years have passed since Bobrow made this claim, and no more sophisticated semantic theory has been forthcoming. Why Bobrow and Minsky think, in the face of the peculiar restrictions necessary to the function of the program, that such a generalization must be possible is hard to understand. Nothing, I think, can justify or even explain their optimism concerning this admittedly limited and ad hoc approach. Their general optimism that some such computable approach must work, how- ever, can be seen to follow from a fundamental metaphysical assumption concerning the nature of language and of human intelligent behavior, namely that whatever orderly behavior people engage in can in principle Phase II (1962-1967) Semantic Information Processing / 49 be formalized and processed by digital computers. (See Chapter 5.) This leads Minsky and Bobrow to shrug off all current difficulties as techno- logical limitations, imposed, for example, by the restricted size of the storage capacity of present machine memories. 25 * Were it not for such an assumption, Bobrow's limited success, her- alded by Minsky as the most promising work thus far, would be recog- nized as a trick which says nothing either for or against the possibility of machine understanding, and the fact that this is the best that an intelligent person like Bobrow could do would lead to discouragement as to the possibility of ever reaching the threshold of self-improving machines. EVANS' ANALOGY PROGRAM The same pattern occurs throughout Minsky's collection: an ad hoc solution of a restricted problem, first reported with caution, and then interpreted as being the first step to more general methods. Yet all the work presented in Minsky's book was completed by 1964, and although seven more years have elapsed, none of the promised generalizations has been produced. Evans* analogy-finding program, for example, is a masterful complex program for solving the sort of analogy problems used in intelligence testing. (See Figure 2.) It performs its particular restricted task as well as an average tenth grader, which, granted the state of the art, is an impressive performance. Evans, moreover, realizes that this success as such has little value unless the techniques he employs can be generalized. But, unlike Bobrow, he does not content himself with the assertion that such a generalization is possible. Rather, he attempts at the end of his paper to sketch the form such a generalization would take, and the contribution it would make to problem-solving programs such as GPS and work in pattern recognition. In the final pages of this chapter we describe a "pattern recognition" process of which the main outlines are based on the conception of ANALOGY described. It is more ambitious chiefly in that a more powerful and more general-purpose descriptive framework for the "objects" is introduced. 26 What Computers Can't Do 750 A is to a as C is to ? Figure 2 GPS treats sub-objects of a given object through its goal-subgoal organization. That is, GPS avoids looking at complex structures on a given level by decompos- ing them into smaller structures tied to subgoals. So GPS never sees a single complex structure as such; when a substructure is handled at some deeper subgoal level it is "out of context" in that the necessary information as to how the achievement of this subgoal contributes to the achievement of larger goals is lacking. Newell discusses a form of this "lack of context" problem and several rather unsatisfactory attempts at solving it. The mechanism we have sketched provides a pattern-recognition device capable of taking a look at the problem which is "global" yet has access to the full structure. Such "global" guidance could be expected to save GPS a large amount of the time now spent in setting up and pursuing subgoals that do not contribute to achieving goals at or near the top level. This alone would be a worthwhile contribution. 27 Evans also has proposals for learning: Certainly the study of these problems in the relatively well-understood domain of phrase-structure languages is a natural next step toward the development of genuine "generalization learning" by machines and a prerequisite to considera- tion of learning in still more complex descriptive language environments. One interesting possibility, since the transformation rules themselves can be described in phrase-structure terms, would be to apply the entire "phrase-structure + GPS" apparatus to improving its own set of transformation rules. 28 Evans realizes that "this may, of course, turn out to be very difficult. 1 ' 29 Presumably it has so turned out, because no more has been published concerning this scheme since this work was completed in 1963, and, as we have seen, since then Newell has abandoned GPS and Murray Eden Phase II (1962-1967) Semantic Information Processing / 57 has reported that in 1968 pattern recognition was as ad hoc as ever. Which, of course, raises the usual question: Why do Minsky and Evans so confidently expect that the ad hoc techniques used to solve this specific and rather complex analogy problem can be generalized? A hint as to the assumptions underlying this confidence can be found in Minsky's surprising comparison of Evans' program to human analogy solving. In spite of his disclaimers that AI is not interested in cognitive simula- tion, Minsky gives the following "mentalistic" description of Evans' pro- gram. To explain the spirit of this work, it is best to describe the program in mentalistic form. Given a set of figures, it constructs a set of hypotheses or theories as follows: 1. Based on the descriptions D(A) and D(B) of Figures A and B [see Figure 2] there are many ways in which D(A) can be transformed into D(B); choose one of these. 2. There are also many ways in which the parts of A can be put into corre- spondence with the parts of C: each such correspondence suggests a rela- tion like that proposed in (1), but which now relates Fig. C and some other figures. 3. It is unlikely that any of the relations found in (2) will apply perfectly to any of the answer-figures. (If just one does, then that will be the program's answer.) For each answer figure, "weaken," i.e., generalize each relation just enough so that it will apply to the figure. 4. Finally, the program measures how much it had to weaken each relation. It chooses the one that required the least change, and gives the correspond- ing answer figure as its answer. By choosing that hypothesis which involved the least "weakening" of the original A >-B transformation hypothesis, the program selects that explanation that contains the most information common to both A >* B and C VD relations. The details of the selection rules in steps (1), (2), (3), and (4), amount, in effect to Evans 9 theory of human behavior in such situations. I feel sure that something of this general character is involved in any kind of analogical rea- soning. 30 This "something" is put more clearly in Minsky's Scientific American article. There he says: "I feel sure that rules or procedures of the same general character are involved in any kind of analogical reasoning." 31 This is the same assumption which, as we have seen, underlies Newell What Computers Can't Do / 52 and Simon's work in CS. In fact, Evans uses a quotation from Newell to describe the problem-solving procedure involved: "These programs are all rather similar in nature. For each the task is difficult enough to allow us to assert that the programs problem-solve, rather than simply carry out the steps of a procedure for a solution invented by the programmer. They all operate on formalized tasks, which, although difficult, are not unstruc- tured. All the programs use the same conceptual approach: they interpret the problem as combinatorial, involving the discovery of the right sequence of opera- tions out of a set of possible sequences. All of the programs generate some sort of tree of possibilities to gradually explore the possible sequences. The set of all sequences is much too large to generate and examine in toto, so that various devices, called heuristics, are employed to narrow the range of possibilities to a set that can be handled within the available limits of processing effort." Evans then concludes: The geometric-analogy program also fits this description. Stated very briefly, given a problem of this type, the program uses various heuristics to select a "correct" rule (in a reasonable time) from a very extensive class of possible rules." It is true that if human beings did solve analogy problems in this way, there would be every reason to expect to be able to improve and general- ize Evans' program, since human beings certainly surpass the machines' present level of performance. But, as in the case of GPS, there is no evidence that human beings proceed in this way, and descriptive, psycho- logical evidence suggests that they do not. Rudolph Arnheim, professor of psychology at Harvard University, in discussing Evans' work, has described the different way in which human beings approach the same sort of problem. His description is worth quoting in full: What happens when a person is confronted with a figure such as Figure [2]? The reaction will vary somewhat from individual to individual as long as no particu- lar context calls for concentration on specific structural features. By and large, however, the observer is likely to notice a vertical arrangement, made up of two units, of which the upper is larger and more complex than the lower; he may also notice a difference in shape. In other words, he will perceive qualitative charac- teristics of placement, relative size, shape; whereas he is unlikely to notice much Phase II (1962-1967) Semantic Information Processing / 53 of the metric properties from which the computer's reading of the pattern must set out, namely, absolute size and the various lengths and distances by which this individual figure is constructed. If one asks observers to copy such a figure, their drawings will show concentration on the topological characteristics and neglect of specific measurements. Confronted now with a pairing of A and B, the human observer may have a rather rich and dazzling experience. He may see, at first, fleeting, elusive resem- blances among basically different patterns. The over-all figure, made up of the pairing of the two, may look unstable, ungraspable, irrational. There are two vertical arrangements, combining in a sort of symmetry; but these two columns are crossed and interfered with by diagonal relations between the two "filled" large circles and the two smaller, unfilled shapes. The various structural features do not add up to a unified, stable, understandable whole. Suddenly, however, the observer may be struck by the simple rectangular arrangement of the four smaller figures: two equal circles on top, two equal squares at the bottom. As soon as this group becomes the dominant theme or structural skeleton of the whole, the remainder the two large circles joins the basic pattern as a sec- ondary, diagonal embellishment. A structural hierarchy has been established. Now the double figure is stable, surveyable, understandable, and therefore ready for comparison with other figures. A first act of problem solving has taken place. If the observer turns to Figure C, his view of this new pattern is determined from the outset by his preceding concern with A and B. Perceived from the viewpoint of A, C reveals a similar vertical structure, distinguished from A mainly by a secondary contrast of shapes. The family resemblance is great, the relation comes easily. But if C is now paired with D t , the resemblance looks excessive, the symmetry too complete. On the contrary, a comparison with D 2 offers too little resemblance. D 3 is recognized immediately as the correct partner, the missing fourth element of the analogy, if the relation between A and B had been properly grasped before. This episode of perceptual problem solving has all the aspects of genuine thinking: the challenge, the productive confusion, the promising leads, the partial solutions, the disturbing contradictions, the flash appearance of a stable solution whose adequacy is self-evident, the structural changes brought about by the pressure of changing total situations, the resemblance discovered among different patterns. It is, in a small way, an exhilarating experience, worthy of a creature endowed with reason; and when the solution has been found, there is a sense of dis-tension, of pleasure, of rest. None of this is true for the computer, not because it is without consciousness, but because it proceeds in a fundamentally different fashion. We are shocked to learn that in order to make the machine solve the analogy problem the experi- What Computers Can't Do / 54 menter "had to develop what is certainly one of the most complex programs ever written." For us the problem is not hard; it is accessible to the brain of a young student. The reason for the difference is that the task calls for the han- dling of topological relations, which require the neglect of purely metric ones. The brain is geared to precisely such topographical features because they inform us of the typical character of things, rather than of their particular measure- ments. 33 As in the case of chess, it turns out that global perceptual grouping is a prior condition for the rule-governed counting out the only kind of procedure available to the machine. As Arnheim puts it, "Topology was discovered by, and relies on, the perceptual powers of the brain, not the arithmetical ones." 34 Obviously Minsky and Evans think that analogies are solved by hu- man beings by applying transformation rules, because the prospects for AI are only encouraging if this is how humans proceed. But it is clearly circular to base one's optimism on an hypothesis which, in turn, is only justified by the fact that if the hypothesis were true, one's optimism would be justified. QUILLIAN'S SEMANTIC MEMORY PROGRAM The final program we shall consider from Phase II, Ross Quillian's Semantic Memory Program, is the most interesting, because most gen- eral; and the most modest, in that its author (working under Simon rather than Minsky) has made no sweeping promises or claims. 35 * This program confirms a general evaluation heuristic already apparent in Samuel's modesty and success and Simon's and Gelernter's claims and setbacks, namely that the value of a program is often inversely propor- tional to its programmer's promises and publicity. Quillian, like Bobrow, is interested in simulating the understanding of natural language; but, unlike Bobrow and Minsky, he sees that this problem cannot be dealt with by ad hoc solutions. In the first place, we do not believe that performance theories or computer models can ignore or put off semantics, as most language processing programs Phase II (1962-1967) Semantic Information Processing / 55 so far have done, and yet hope to achieve success. Whether a program is intended to parse sentences, to translate languages, or to answer natural language ques- tions, if it does not take account of semantic facts both early and often, I do not think it has a chance of approaching the level of human competence. 36 After reviewing all work in the field, including that of Bobrow, Quil- lian remarks: Programs such as Bobrow's have been able to set up the equations corresponding to certain algebra word problems by an almost entirely "syntactic" procedure. . . . However, if one attempts to extend the range of language that such a program can handle, it becomes necessary to incorporate increasing numbers of semantic facts. 37 Quillian concludes that the problems of what is to be contained in an overall, human-like permanent memory, what format this is to be in, and how this memory is to be organized have not been dealt with in great generality in prior simulation programs. . . . Further advances in simulating problem-solving and game playing, as well as language performance, will surely require programs that develop and interact with large memories. 38 Quillian then proceeds to propose a complex heuristic program for storing and accessing the meaning of words and "anything that can be stated in language, sensed in perception, or otherwise known and remem- bered" 39 in one "enormous interlinked net." 40 Quillian proposes this program as "a reasonable view of how semantic information is organized within a person's memory." 41 He gives no argument to show that it is reasonable except that if a computer were to store semantic information, this would be a reasonable model for it. People, indeed, are not aware of going through any of the complex storage and retrieval process Quil- lian outlines, but this does not disturb Quillian, who, like his teacher, Simon, in similar trouble can always claim that these processes are nonetheless unconsciously taking place: While the encoding process is of course not identical to the covert processing that constitutes the understanding of the same text during normal reading, it is ... in some ways a slowed-down, overt version of it. 42 What Computers Can't Do / 56 That such unconscious processing is going on, and moreover, that such processing follows heuristic rules is by no means obvious. We have seen in the cases of chess playing and analogy solving that gestalt group- ing plays a crucial role, and it may well do so here. Yet Quillian seems to have inherited Newell and Simon's unquestioned assumption that human beings operate by heuristic programs. The heuristic methods by which one particular comprehension of text is selected is the central problem for anyone who would explain "understanding," just as the heuristic methods by which one particular chess move is selected from all those possible is the central problem for anyone who would explain chess playing. 43 In terms of this assumption Quillian must assume that the task of the program involves working from parts to wholes. In selecting a task to perform with a model memory, one thinks first of the ability to understand unfamiliar sentences. It seems reasonable to suppose that people must necessarily understand new sentences by retrieving stored information about the meaning of isolated words and phrases, and then combining and perhaps altering these retrieved word meanings to build up the meanings of sentences. Accordingly, one should be able to take a model of stored semantic knowledge, and formulate rules of combination that would describe how sen- tence meanings get built up from stored word meanings. 44 Quillian also has great hopes for his system: It further seems likely that if one could manage to get even a few word meanings adequately encoded and stored in a computer memory, and a workable set of combination rules formalized as a computer program, he could then bootstrap his store of encoded word meanings by having the computer itself "understand" sentences that he had written to constitute the definitions of other single words. That is, whenever a new, as yet uncoded, word could be defined by a sentence using only words whose meanings had already been encoded, then the represen- tation of this sentence's meaning, which the machine could build by using its previous knowledge together with its combination rules, would be the ap- propriate representation to add to its memory as the meaning of the new word. 43 But with a frankness, rare in the literature, Quillian also reports his disappointments: Phase II (1962-1967) Semantic Information Processing / 57 Unfortunately, two years of work on this problem led to the conclusion that the task is much too difficult to execute at our present stage of knowledge. The processing that goes on in a person's head when he "understands" a sentence and incorporates its meaning into his memory is very large indeed, practically all of it being done without his conscious knowledge. 46 The magnitude of the problem confronting Quillian becomes clear when we note that the definition of eight hundred and fifty words comprise far more information than can be modeled in the core of today's computers. . . . 47 These difficulties suggest that the model itself the idea that our un- derstanding of a natural language involves building up a structured whole out of an enormous number of explicit parts may well be mis- taken. Quillian's work raises rather than resolves the question of storing the gigantic number of facts resulting from an analysis which has no place for perceptual gestalts. If this data structure grows too rapidly with the addition of new definitions, then Quillian's work, far from being encouraging, would be a reductio ad absurdum of the whole computer- oriented approach. Before taking a stand on whether Quillian's work is grounds for optimism, one would expect an answer to the basic question: Does Quillian's data base grow linearly or exponentially with additional entries? On this crucial point it is surprising to find much hope but little information. Quillian's program contains definitions of only from 50 to 60 words, and, in describing Quillian's work, in his book written in 1968, three years after the work was completed, Minsky has to admit that "we simply do not know enough about how powerful Quillian's methods would be when provided with a more substantial knowledge bank." 48 Again, no further progress has been reported. II. Significance of Current Difficulties What would be reasonable to expect? Minsky estimates that Quillian's program now contains a few hundred facts. He estimates that "a million facts would be necessary for great intelligence." 49 He also admits that What Computers Can't Do / 58 each of "the programs described [in this book] will work best when given exactly the necessary facts, and will bog down inexorably as the informa- tion files grow." 50 Is there, thus, any reason to be confident that these programs are approaching the "superior heuristics for managing their knowledge structure" which Minsky believed human beings must have; or, as Minsky claims in another of his books, that within a generation . . . few compartments of intellect will remain outside the machine's realm the problem of creating "artificial intelligence" will be sub- stantially solved. 51 Certainly there is nothing in Semantic Information Processing to justify this confidence. As we have seen, Minsky criticizes the early programs for their lack of generality. "Each program worked only on its restricted specialty, and there was no way to combine two different problem- solvers." 52 But Minsky's solutions are as ad hoc as ever. Yet he adds jauntily: The programs described in this volume may still have this character, but they are no longer ignoring the problem. In fact, their chief concern is finding methods of solving it. 53 But there is no sign that any of the papers presented by Minsky have solved anything. They have not discovered any general feature of the human ability to behave intelligently. All Minsky presents are clever special solutions, like Bobrow's and Evans', or radically simplified mod- els such as Quillian's, which work because the real problem, the problem of how to structure and store the mass of data required has been put aside. Minsky, of course, has already responded to this apparent short- coming with a new version of the first step fallacy: The fact that the present batch of programs still appear to have narrow ranges of application does not indicate lack of progress toward generality. These pro- grams are steps toward ways to handle knowledge. 54 In Phase II the game seems to be to see how far one can get with the appearance of complexity before the real problem of complexity has to Phase II (1962-1967) Semantic Information Processing / 59 be faced, and then when one fails to generalize, claim to have made a first step. Such an approach is inevitable as long as workers in the field of AI are interested in producing striking results but have not solved the practical problem of how to store and access the large body of data necessary, if perhaps not sufficient, for full-scale, flexible, semantic infor- mation processing. Minsky notes with satisfaction, looking over the results, "one cannot help being astonished at how far they did get with their feeble semantic endowment." 55 Bar-Hillel in a recent talk to SI- GART (Special Interest Group in Artificial Intelligence of the Associa- tion for Computing Machinery) calls attention to the misleading character of this sort of claim. There are very many people in all fields but particularly in the field of AI who, whenever they themselves make a first step towards letting the computer do certain things it has never done before, then believe that the remaining steps are nothing more than an exercise in technical ability. Essentially, this is like saying that as soon as anything can be done at all by a computer, it can also be done well. On the contrary, the step from not being able to do something at all to being able to do it a little bit is very much smaller than the next step being able to do it well. In AI, this fallacious thinking seems to be all perva- sive. 56 But Bar-Hillel is too generous in suggesting that the fallacy is sim- ply overestimation of the ease of progress. To claim to have taken even an easy first step one must have reason to believe that by further such steps one could eventually reach one's goal. We have seen that Min- sky's book provides no such reasons. In fact these steps may well be strides in the opposite direction. The restricted character of the results reported by Minsky, plus the fact that during the last five years none of the promised generalizations has been produced, suggests that human beings do not deal with a mass of isolated facts as does a digital computer, and thus do not have to store and retrieve these facts by heuristic rules. Judging from their behavior, human beings avoid rather than resolve the difficulties confronting workers in Cognitive Simu- lation and Artificial Intelligence by avoiding the discrete informa- What Computers Can't Do / 60 tion-processing techniques from which these difficulties arise. Thus it is by no means obvious that Minsky's progress toward handling "knowledge" (slight as it is) is progress toward artificial intelligence at all. Conclusion We have seen how Phase I, heralded as a first step, ends with the abandonment of GPS and the general failure to provide the theorem proving, chess playing, and language translation programs anticipated. Minsky himself recognizes this failure, and while trying to minimize it, diagnoses it accurately: A few projects have not progressed nearly as much as was hoped, notably projects in language translation and mathematical theorem-proving. Both cases, I think, represent premature attempts to handle complex formalisms without also representing their meaning. 1 Phase II a new first step begins around 1961 with Minsky 's gradu- ate students at M.I.T. undertaking theses aimed at overcoming this difficulty. It ends in 1968 with the publication of Minsky's book Semantic Information Processing, which reports these attempts, all completed by 1964. After analyzing the admittedly ad hoc character of those programs which Minsky considers most successful, and noting the lack of follow- up during the last five years, we can only conclude that Phase II has also ended in failure. Most reports on the state of the art try to cover up this failure. In a report undertaken for the IEEE in 1966, covering work in AI since 1960, R. J. Solomonoff devotes his first three pages to speaking of GPS and other past achievements, already completed by 1960, and the next three What Computers Can't Do / 62 pages to talking of the glorious future of work on induction by S. Amarel: "Although Amarel hasn't programmed any of his theories, his ideas and his analysis of them are important." 2 There is little mention of the semantic information processing programs touted by Minsky. All hope is placed in induction and learning. Unfortunately, "in all the learning systems mentioned, the kinds of self improvement accessible to the ma- chines have been quite limited. . . . We still need to know the kind of heuristics we need to find heuristics, as well as what languages can readily describe them." 3 Since no one has made any contribution to finding these heuristics, Solomonoff s final hope is placed in artificial evolution: The promise of artificial evolution is that many things are known or suspected about the mechanisms of natural evolution, and that those mechanisms can be used directly or indirectly to solve problems in their artificial counterparts. For artificial intelligence research, simulation of evolution is incomparably more promising than simulation of neural nets, since we know practically nothing about natural neural nets that would be at all useful in solving difficult problems. 4 This work in artificial evolution, however, has hardly begun. "Research in simulation of evolution has been very limited in both quantity and quality." 5 When an article supposed to sum up work done since 1960 begins with earlier accomplishments and ends in speculations, without presenting a single example of actual progress, stagnation can be read between the lines. Occasionally one catches hints of disappointment in the lines them- selves. For example, Fred Tonge, whose solid, unpretentious paper on a heuristic line-balancing procedure was reprinted in Computers and Thought, after reviewing progress in AI, concluded in 1968: While many interesting programs (and some interesting devices) have been pro- duced, progress in artificial intelligence has not been exciting or spectacular. . . . This is due at least in part to lack of a clear separation between accomplish- ment and conjecture in many past and current writings. In this field, as in many others, there is a large difference between saying that some accomplishment "ought to" be possible and doing it. Identifiable, significant, landmarks of accomplishment are scarce. 6 Conclusion / 63 Tonge then gives his list of "landmarks." They are Newell, Shaw, and Simon's Logic Theory Machine, Samuel's Checker Program, and the Uhr-Vossler pattern recognition program all completed long before 1961, and all dead ends if we are to judge from subsequent work. That mine is no unduly prejudiced reaction to Tonge's summary of the work done thus far can be seen by comparing P. E. Greenwood's review of Tonge's article for Computing Reviews: "From this brief summary of the state of the art of artificial intelligence, one would conclude that little significant progress has been made since about 1960 and the prospects for the near future are not bright." 7 Why, in the light of these difficulties, do those pursuing Cognitive Simulation assume that the information processes of a computer reveal the hidden information processes of a human being, and why do those working in Artificial Intelligence assume that there must be a digital way of performing human tasks? To my knowledge, no one in the field seems to have asked himself these questions. In fact, artificial intelligence is the least self-critical field on the scientific scene. There must be a reason why these intelligent men almost unanimously mimimize or fail to recognize their difficulties, and continue dogmatically to assert their faith in prog- ress. Some force in their assumptions, clearly not their success, must allow them to ignore the need for justification. We must now try to discover why, in the face of increasing difficulties, workers in artificial intelligence show such untroubled confidence. PARTI ASSUMPTIONS UNDERLYING PERSISTENT OPTIMISM Introduction In spite of grave difficulties, workers in Cognitive Simulation and Artificial Intelligence are not discouraged. In fact, they are unqualifiedly optimistic. Underlying their optimism is the conviction that human information processing must proceed by discrete steps like those of a digital computer, and, since nature has produced intelligent behavior with this form of processing, proper programming should be able to elicit such behavior from digital machines, either by imitating nature or by out-programming her. This assumption, that human and mechanical information processing ultimately involves the same elementary processes, is sometimes made naTvely explicit. Newell and Simon introduce one of their papers with the following remark: It can be seen that this approach makes no assumption that the "hardware** of computers and brains are similar, beyond the assumptions that both are general- purpose symbol-manipulating devices, and that the computer can be pro- grammed to execute elementary information processes functionally quite like those executed by the brain. 1 But this is no- innocent and empty assumption. What is a general- purpose symbol-manipulating device? What are these "elementary infor- mation processes" allegedly shared by man and machine? All artificial intelligence work is done on digital computers because they are the only What Computers Can't Do / 68 general-purpose information-processing devices which we know how to design or even conceive of at present. All information with which these computers operate must be represented in terms of discrete elements. In the case of present computers the information is represented by binary digits, that is, in terms of a series of yeses and noes, of switches being open or closed. The machine must operate on finite strings of these determinate elements as a series of objects related to each other only by rules. Thus the assumption that man functions like a general-purpose symbol-manipulating device amounts to 1. A biological assumption that on some level of operation usually supposed to be that of the neurons the brain processes information in discrete operations by way of some biological equivalent of on/off switches. 2. A psychological assumption that the mind can be viewed as a device operating on bits of information according to formal rules. Thus, in psychology, the computer serves as a model of the mind as conceived of by empiricists such as Hume (with the bits as atomic impressions) and idealists such as Kant (with the program providing the rules). Both empiricists and idealists have prepared the ground for this model of thinking as data processing a third-person process in which the in- volvement of the "processor" plays no essential role. 3. An epistemological assumption that all knowledge can be formal- ized, that is, that whatever can be understood can be expressed in terms of logical relations, more exactly in terms of Boolean functions, the logical calculus which governs the way the bits are related according to rules. 4. Finally, since all information fed into digital computers must be in bits, the computer model of the mind presupposes that all relevant information about the world, everything essential to the production of intelligent behavior, must in principle be analyzable as a set of situation- free determinate elements. This is the ontological assumption that what there is, is a set of facts each logically independent of all the others. In the following chapters we shall turn to an analysis of the plausibility of each of these assumptions. In each case we shall see that the assump- Introduction / 69 tion is taken by workers in CS or AI as an axiom, guaranteeing results, whereas it is, in fact, only one possible hypothesis among others, to be tested by the success of such work. Furthermore, none of the four assumptions is justified on the basis of the empirical and a priori argu- ments brought forward in its favor. Finally, the last three assumptions, which are philosophical rather than empirical, can be criticized on phil- osophical grounds. They each lead to conceptual difficulties when fol- lowed through consistently as an account of intelligent behavior. After we have examined each of these assumptions we shall be in a better position to understand the persistent optimism of workers in artificial intelligence and also to assess the true significance of results obtained thus far. The Biological Assumption In the period between the invention of the telephone relay and its apotheosis in the digital computer, the brain, always understood in terms of the latest technological inventions, was understood as a large tele- phone switchboard or, more recently, as an electronic computer. This model of the brain was correlated with work in neurophysiology which found that neurons fired a somewhat all-or-nothing burst of electricity. This burst, or spike, was taken to be the unit of information in the brain corresponding to the bit of information in a computer. This model is still uncritically accepted by practically everyone not directly involved with work in neurophysiology, and underlies the naive assumption that man is a walking example of -a successful digital- computer program. But to begin with, even if the brain did function like a digital computer at some level it would not necessarily provide encouragement for those working in CS or AI. For the brain might be wired like a very large array of randomly connected neurons, such as the perceptrons proposed by the group Minsky dismisses as the early cyberneticists. 1 * Such a neural net can be simulated using a program, but such a program is in no sense a heuristic program. Thus the mere fact that the brain might be a digital computer is in no way ground for optimism as to the success of artificial intelligence as defined by Simon or Minsky. Moreover, it is an empirical question whether the elementary informa- What Computers Can't Do / 72 tion processing in the brain can best be understood in terms of a digital model. The brain might process information in an entirely different way than a digital computer does. Information might, for example, be pro- cessed globally the way a resistor analogue solves the problem of the minimal path through a network. Indeed, current evidence suggests that the neuron-switch model of the brain is no longer empirically tenable. Already in 1956 John von Neumann, one of the inventors of the modern digital computer, had his doubts: Now, speaking specifically of the human nervous system, this is an enormous mechanism at least 10 6 times larger than any artifact with which we are familiar and its activities are correspondingly varied and complex. Its duties include the interpretation of external sensory stimuli, of reports of physical and chemical conditions, the control of motor activities and of internal chemical levels, the memory function with its very complicated procedures for the transformation of and the search for information, and of course, the continuous relaying of coded orders and of more or less quantitative messages. It is possible to handle all these processes by digital methods (i.e., by using numbers and expressing them in the binary system or, with some additional coding tricks, in the decimal or some other system), and to process the digitalized, and usually numericized, informa- tion by algebraical (i.e., basically arithmetical) methods. This is probably the way a human designer would at present approach such a problem. The available evidence, though scanty and inadequate, rather tends to indicate that the human nervous system uses different principles and procedures. Thus message pulse trains seem to convey meaning by certain analogic traits (within the pulse notation i.e., this seems to be a mixed, part digital, part analog system), like the time density of pulses in one line, correlations of the pulse time series between different lines in a bundle, etc. 2 Von Neumann goes on to spell out what he takes to be the "mixed character of living organisms." The neuron transmits an impulse. . . . The nerve impulse seems in the main to be an all-or-none affair, comparable to a binary digit. Thus a digital element is evidently present, but it is equally evident that this is not the entire story. ... It is well known that there are various composite functional sequences in the organism which have to go through a variety of steps from the original stimulus to the ultimate effect some of the steps being neural, that is, digital, and others humoral, that is, analog. 3 The Biological Assumption / 73 But even this description grants too much to the digital model. It does not follow from the fact that the nerve impulse is an all-or-none affair that any digital processing at all is taking place. The distinction between digital and analogue computation is a logical distinction, not a distinc- tion based on the hardware or the sort of electrical impulses in the system. The essential difference between digital and analogue informa- tion processing is that in digital processing a single element represents a symbol in a descriptive language, that is, carries a specific bit of information; while in a device functioning as an analogue computer, continuous physical variables represent the information being processed. The brain, operating with volleys of pulses, would be a digital computer only if each pulse were correlated with some step in an information- processing sequence; if, however, the rate at which pulses are transmitted turns out to be the minimum unit which can be assigned a place in an information-processing model as von Neumann seems to hold then the brain would be operating as an analogue device. 4 * Once this conceptual confusion has been cleared up, von Neumann can be understood as suggesting that the brain functions exclusively like an analogue computer, and subsequent work has tended to confirm this hypothesis. Even for those unfamiliar with the technical details of the following report, the conclusion is clear: In the higher invertebrates we encounter for the first time phenomena such as the graded synaptic potential, which before any post-synaptic impulse has arisen can algebraically add the several incoming presynaptic barrages in a complex way. These incoming barrages are of a different value depending upon the pathway and a standing bias. Indeed, so much can be done by means of this graded and nonlinear local phenomenon prior to the initiation of any post- synaptic impulse that we can no more think of the typical synapse in integrative systems as being a digital device exclusively as was commonly assumed a few years ago, but rather as being a complex analog device. . . . 5 The latest suggestion from Jerome Lettvin of M.I.T. is that the diame- ter of the axion may play a crucial role in processing information by acting as a filter. 6 An individual neuron fires at a certain frequency. The diameter of its various axion branches would act as low pass filters at What Computers Can't Do / 74 different cutoff frequencies. Output from a given cell would then produce different frequencies at different terminals. The filter characteristics of the axion would vary with its diameter which in turn might be a function of the recency of signals passing down that axion, or even, perhaps, of the activation of immediately environing axions. If such time factors and field interactions play a crucial role, there is no reason to hope that the information processing on the neurophysiological level can be described in a digital formalism or, indeed, in any formalism at all. In 1966, Walter Rosenblith of M.I.T., one of the pioneers in the use of computers in neuropsychology, summed up the situation: We no longer hold the earlier widespread belief that the so-called all-or-none law from nerve impulses makes it legitimate to think of relays as adequate models for neurons. In addition, we have become increasingly impressed with the in- teractions that take place among neurons: in some instances a sequence of nerve impulses may reflect the activities of literally thousands of neurons in a finely graded manner. In a system whose numerous elements interact so strongly with each other, the functioning of the system is not necessarily best understood by proceeding on a neuron-by-neuron basis as if each had an independent personal- ity. . . . Detailed comparisons of the organization of computer systems and brains would prove equally frustrating and inconclusive. 7 Thus the view that the brain as a general-purpose symbol-manipulat- ing device operates like a digital computer is an empirical hypothesis which has had its day. No arguments as to the possibility of artificial intelligence can be drawn from current empirical evidence concerning the brain. In fact, the difference between the "strongly interactive" nature of brain organization and the noninteractive character of machine organization suggests that insofar as arguments from biology are rele- vant, the evidence is against the possibility of using digital computers to produce intelligence. The Psychological Assumption Whether the brain operates like a digital computer is a strictly empiri- cal question to be settled by neurophysiology. The computer model simply fails to square with the facts. No such simple answer can be given to the related but quite different question: whether the mind functions like a digital computer, that is, whether one is justified in using a com- puter model in psychology. The issue here is much harder to define. The brain is clearly a physical object which uses physical processes to trans- form energy from the physical world. But if psychology is to differ from biology, the psychologist must be able to describe some level of function- ing other than the physical-chemical reactions in the brain. The theory we shall criticize claims that there is such a level the information-processing level and that on this level the mind uses com- puter processes such as comparing, classifying, searching lists, and so forth, to produce intelligent behavior. This mental level, unlike the physi- cal level, has to be introduced as a possible level of discourse. The issues involved in this discussion will, therefore, be philosophical rather than empirical. We shall see that the assumption of an information-processing level is by no means so self-evident as the cognitive simulators seem to think; that there are good reasons to doubt that there is any "information processing" going on, and therefore reason to doubt the validity of the claim that the mind functions like a digital computer. What Computers Can't Do / 76 In 1957 Simon predicted that within ten years psychological theories would take the form of computer programs, and he set out to fulfill this prediction by writing a series of programs which were meant to simulate human cognition by simulating the conscious and unconscious steps a person goes through to arrive at a specific cognitive performance. And we have seen that despite the general inadequacy of such programs, admitted even by enthusiasts such as Minsky, all those involved in the general area of artificial intelligence (Minsky included) share the as- sumption that human beings, when behaving intelligently, are following heuristic rules similar to those which would be necessary to enable a digital computer to produce the same behavior. Moreover, despite meager results, Simon's prediction has nonetheless been partially fulfilled. There has been a general swing from behaviorism to mentalism in psychology. Many influential psychologists and philoso- phers of psychology have jumped on Simon's bandwagon and begun to pose their problems in terms of computer analogies. Ulric Neisser as- sumes that "the task of a psychologist trying to understand human cognition is analogous to that of a man trying to discover how a com- puter has been programmed." 1 And George Miller of Harvard now speaks of "recent developments in our understanding of man viewed as a system for processing information." 2 Usually no argument is given for this new dogma that man is an information-processing system functioning like a heuristically pro- grammed digital computer. It seems rather to be an unquestioned axiom underlying otherwise careful and critical analysis. There is no doubt some temptation to suppose that since the brain is a physical thing and can be metaphorically described as "processing information," there must be an information-processing level, a sort of flow chart of its operations, in which its information-processing activity can be described. But we have seen in Chapter 3 that just because the brain is physical and pro- cesses information is no reason for biologists to suppose that it functions like a digital computer. The same holds for the psychological level. Although psychologists describe that function called the mind as "pro- cessing information," this does not mean that it actually processes infor- The Psychological Assumption / 77 mation in the modern technical sense, nor that it functions like a digital computer, that is, that it has a program. "Information processing" is ambiguous. If this term simply means that the mind takes account of meaningful data and transforms them into other meaningful data, this is certainly incontrovertible. But the cyber- netic theory of information, introduced in 1948 by Claude Shannon, has nothing to do with meaning in this ordinary sense. It is a nonsemantic, mathematical theory of the capacity of communication channels to transmit data. A bit (binary digit) of information tells the receiver which of two equally probable alternatives has been chosen. In his classic paper "The Mathematical Theory of Communication" Shannon was perfectly clear that his theory, worked out for telephone engineering, carefully excludes as irrelevant the meaning of what is being transmitted. The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem. 3 Warren Weaver in explaining the significance of Shannon's paper is even more emphatic: The word Information, in this theory, is used in a special sense that must not be confused with its ordinary usage. In particular, information must not be confused with meaning. In fact, two messages, one of which is heavily loaded with meaning and the other of which is pure nonsense, can be exactly equivalent, from the present viewpoint, as regards information. It is this, undoubtedly, that Shannon means when he says that "the semantic aspects of communication are irrelevant to the engineering aspects." 4 When illegitimately transformed into a theory of meaning, in spite of Shannon's warning, information theory and its vocabulary have already built in the computer-influenced assumption that experience can be analyzed into isolable, atomic, alternative choices. As a theory of mean- ing this assumption is by no means obvious. Gestalt psychologists, for example, claim (as we have seen in Part I and shall argue in detail in Part What Computers Can't Do / 78 III) that thinking and perception involve global processes which cannot be understood in terms of a sequence or even a parallel set of discrete operations. 5 * Just as the brain seems to be, at least in part, an analogue computer, so the mind may well arrive at its thoughts and perceptions by responding to "fields," "force," "configurations," and so on, as, in fact, we seem to do insofar as our thinking is open to phenomenological description. 6 It is precisely the role of the programmer to make the transition from statements which are meaningful (contain information in the ordinary sense) to the strings of meaningless discrete bits (information in the technical sense) with which a computer operates. The ambition of artifi- cial intelligence is to program the computer to do this translating job itself. But it is by no means obvious that the human translator can be dispensed with. Much of the literature of Cognitive Simulation gains its plausibility by shifting between the ordinary use of the term "information" and the special technical sense the term has recently acquired. Philosophical clarity demands that we do not foreclose the basic question whether human intelligence presupposes rulelike operations on discrete elements before we begin our analysis. Thus we must be careful to speak and think of "information processing" in quotation marks when referring to hu- man beings. Moreover, even if the mind did process information in Shannon's sense of the term, and thus function like a digital computer, there is no reason to suppose that it need do so according to a program. If the brain were a network of randomly connected neurons, there might be no flow chart, no series of rule-governed steps on the information-processing level, which would describe its activity. Both these confusions the step from ordinary meaning to the techni- cal sense of information and from computer to heuristically programmed digital computer are involved in the fallacy of moving from the fact that the brain in some sense transforms its inputs to the conclusion that the brain or mind performs some sequence of discrete operations. This fallacy is exhibited in the baldest form in a recent paper by Jerry Fodor. It is instructive to follow his argument. The Psychological Assumption / 79 Fodor begins with generally accepted facts about the central nervous system: If the story about the causal determination of depth estimates by texture gradi- ents is true and if the central nervous system is the kind of organ most sensitive people now think it is, then some of the things the central nervous system does, some of the physical transactions that take place in the central nervous system when we make estimates of depth, must satisfy such descriptions as 'monitoring texture gradients', 'processing information about texture gradients', 'computing derivatives of texture gradients', etc. 7 He thus arrives at the view that "every operation of the nervous system is identical with some sequence of elementary operations." 8 Disregarding the question-begging use of "processing information" in this account, we can still object that computing the first derivative of a texture gradient is the sort of operation very likely to be performed by some sort of analogue device. There is, therefore, no reason at all to conclude from the fact that the nervous system responds to differences in texture gradients that "every operation of the nervous system is identi- cal with some sequence of elementary operations. . . ." There is, indeed, not the slightest justification for the claim that "for each type of behavior in the repertoire of that organism, a putative answer to the question, How does one produce behavior of that type? takes the form of a set of specific instructions for producing the behavior by performing a set of machine operations." 9 The argument gains its plausibility from the fact that if a psychologist were to take the first derivative of a texture gradient, he would compute it using a formalism (differential calculus) which can be manipulated in a series of discrete operations on a digital computer. But to say that the brain is necessarily going through a series of operations when it takes the texture gradient is as absurd as to say that the planets are necessarily solving differential equations when they stay in their orbits around the sun, or that a slide rule (an analogue computer) goes through the same steps when computing a square root as does a digital computer when using the binary system to compute the same number. Consider an ion solution which might be capable of taking a texture gradient or of simulating some other perceptual process by reaching What Computers Can't Do / 80 equilibrium. Does the solution, in reaching equilibrium, go through the series of discrete steps a digital computer would follow in solving the equations which describe this process? In that case, the solution is solv- ing in moments a problem which it would take a machine centuries to solve if the machine could solve it at all. Is the solution an ultrarapid computer, or has it got some supposedly clever heuristic like the chess master, which simplifies the problem? Obviously, neither. The fact that we can describe the process of reaching equilibrium in terms of equations and then break up these equations into discrete elements in order to solve them on a computer does not show that equilibrium is actually reached in discrete steps. Likewise, we need not conclude from the fact that all continuous physicochemical processes involved in human "information processing" can in principle be formalized and calculated out discretely, that any discrete processes are actually taking place. Moreover, even if one could write such a computer program for simu- lating the physicochemical processes in the brain, it would be no help to psychology. If simulation is taken in its weakest possible sense, a device is simu- lated by any program which realizes the same input/output function (within the range of interest). Whether achievable for the brain or not, this clearly lacks what is necessary for a psychological theory, namely an account of how the mind actually "works." For psychological expla- nation, a representation, somehow stronger than a mere simulation, is required. As Fodor notes: We can say that a machine is strongly equivalent to an organism in some respect when it is weakly equivalent in that same respect and the processes upon which the behavior of the machine is contingent are of the same type as the processes upon which the behavior of the organism are contingent. 10 That is, equivalence in the psychological respect demands machine pro- cesses, of the psychological type. n * Psychological operations must be the sort which human beings at least sometimes consciously perform when processing information for example, searching, sorting, and storing and not physicochemical processes in the organism. Thus a chess player's The Psychological Assumption / 57 report as he zeroed in on his Rook, "And now my brain reaches the following chemical equilibrium, described by the following array of diff- erential equations," would describe physiological processes no doubt correlated with "information processing," but not that "information processing" itself. Fodor is not clear whether his argument is supposed to be a priori or empirical, that is, whether or not he thinks it follows logically or merely contingently from the claim that the brain is taking account of the texture gradient that it is performing a sequence of elementary opera- tions. The fact that he chooses this example, which is one of the least plausible cases in which one would want to argue that the brain or the mind is performing any elementary operations at all, suggests that he thinks there is some kind of necessary connection between taking a texture gradient, computing, and performing a sequence of operations. When this argument is shown to be a series of confusions, however, the advocates of the psychological assumption can always shift ground and claim that theirs is not an a priori argument but an empirical conclusion based on their experiments. Fodor took this tack in defending his paper at the meeting of the American Philosophical Association at which it was delivered, while Miller et al. justify their work strictly on the basis of what they take to be the success of CS. A Plan is, for an organism, essentially the same as a program for a computer. . . . Newell, Shaw, and Simon have explicitly and systematically used the hier- archical structure of lists in their development of "information-processing lan- guages" that are used to program high-speed digital computers to simulate human thought processes. Their success in this direction which the present authors find most impressive and encouraging argues strongly for the hypothesis that a hierarchical structure Is the basic form of organization of human problem-solving. 12 We have seen in Part I that Newell, Shaw, and Simon's results are far from impressive. What then is this encouraging empirical evidence? We must now look at the way Newell, Shaw, and Simon's work is evaluated. What Computers Can't Do / 82 I. Empirical Evidence for the Psychological Assumption: Critique of the Scientific Methodology of Cognitive Simulation The empirical justification of the psychological assumption poses a ques- tion of scientific methodology the problem of the evaluation of evi- dence. Gross similarities of behavior between computers and people do not justify the psychological asumption, nor does the present inability to demonstrate these similarities in detail alone justify its rejection. A test of the psychological assumption requires a detailed comparison of the steps involved in human and machine information processing. As we have seen (Chapter 1, Sec. II), Newell, Shaw, and Simon conscientiously note the similarities and differences between human protocols and ma- chine traces recorded during the solution of the same problem. We must now turn to their evaluation of the evidence thus obtained. Newell and Simon conclude that their work provide[s] a general framework for understanding problem-solving behavior . . . and finally reveals with great clarity that free behavior of a reasonably intelligent human can be understood as the product of a complex but finite and determinate set of laws. 13 This is a strangely unscientific conclusion to draw, for Newell and Simon acknowledge that their specific theories like any scientific theories must stand or fall on the basis of their generality, that is, the range of phenomena which can be explained by the programs. 14 Yet their program is nongeneral in at least three ways. The available evidence has neces- sarily been restricted to those most favorable cases where the subject can to some extent articulate his information-processing protocols (game playing and the solution of simple problems) to the exclusion of pattern recognition and the acquisition and use of natural language. Moreover, even in these restricted areas the machine trace can only match the performance of one individual, and only after ad hoc adjustments. And finally, even the match is only partial. Newell and Simon note that their program "provides a complete explanation of the subject's task behavior The Psychological Assumption / 83 with five exceptions of varying degrees of seriousness." 15 In the light of these restrictions it is puzzling how Newell and Simon can claim a "general framework," and in the light of the exceptions it is hard to see how they can claim to have any kind of scientific under- standing at all. There seems to be some confusion here concerning the universality of scientific laws or theories. In general, scientific laws do not admit of exceptions, yet here the exceptions are honestly noted as if the frank recognition of these exceptions mitigates their importance; as if Galileo might, for example, have presented the law of falling bodies as holding for all but five objects which were found to fall at a different rate. Not that a scientific conjecture must necessarily be discarded in the face of a few exceptions; there are scientifically sanctioned ways of deal- ing with such difficulties. One can, to begin with, hold on to the generali- zation as a working hypothesis and wait to announce a scientific law until the exceptions are cleared up. A working hypothesis need not explain all the data. When, however, the scientist claims to present a theory, let alone a "general framework for understanding," he must deal with these exceptions either by subsuming them under the theory (as in the appeal to friction to explain deviations from the laws of motion), or by suggest- ing where to look for an explanation, or at least by showing how, accord- ing to the theory, one would expect such difficulties. Newell and Simon take none of these lines. They might argue that there is no cause for concern, that there are exceptions to even the best theories. In his study of scientific revolutions, Thomas Kuhn notes the persistence of anomalies in all normal science. There are always some discrepancies. . . . Even the most stubborn ones usually respond at last to normal practice. Very often scientists are willing to wait, particularly if there are many problems available in other parts of the field. We have already noted, for example, that for sixty years after Newton's original computation, the predicted motion of the moon's perigee remained only half of that observed. 16 But this cannot be a source of comfort for Newell and Simon. Such tolerance of anomalies assumes that there already is an ongoing science, an "accepted paradigm" which "must seem better than its competi- What Computers Can't Do / 84 tors." 17 This supposes that the theory works perfectly in at least some clearly defined area. But Newell and Simon's cognitive theory is not only not general. It does not work even in a carefully selected special case. It is just where we would have to find a perfect match in order to establish a paradigm that we find the exceptions. Thus Newell and Simon's work, even though it offers some surprising approximations, does not establish a functioning science which would justify a claim to have found general laws even in the face of anomalies. In discussing the Newtonian anomaly above, Kuhn goes on to point out that "Europe's best mathematical physicists continued to wrestle unsuccessfully with the well-known discrepancy. . . ." l8 The absence of this sort of concern further distinguishes Newell and Simon's work from normal scientific practice. After noting their exceptions, no one in CS least of all Newell and Simon seems interested in trying to account for them. Rather all go on to formulate, in some new area, further ad hoc rough generalizations. There is one other acceptable way of dealing with exceptions. If one knew, on independent grounds, that mental processes must be the prod- uct of a rule-governed sequence of discrete operations, then exceptions could be dealt with as accidental difficulties in the experimental tech- nique, or challenging cases still to be subsumed under the law. Only then would those involved in the field have a right to call each program which simulated intelligent behavior no matter how approximately an achievement and to consider all setbacks nothing but challenges for sharper heuristic hunting and further programming ingenuity. The prob- lem, then, is how to justify independently the assumption that all human "information processing" proceeds by discrete steps. Otherwise the ex- ceptions, along with the narrow range of application of the programs and the lack of progress during the last ten years, would tend to disconfirm rather than confirm this hypothesis. The "justification" seems to have two stages. In the early literature, instead of attempting to justify this important and questionable digital-assumption, Newell and Simon present it as a postulate, a working hypothesis which directs their investigation. "We The Psychological Assumption / 85 postulate that the subject's behavior is governed by a program organized from a set of elementary information processes." 19 This postulate, which alone might seem rather arbitrary, is in turn sanctioned by the basic methodological principle of parsimony. According to Newell, Shaw, and Simon, this principle enjoins us to assume tentatively the most simple hypothesis, in this case that all information processing resembles that sort of processing which can be programmed on a digital computer. We can suppose, for example, that in chess, when our subject is zeroing in, he is unconsciously counting out. In general, whenever the machine trace shows steps which the subject did not report, the principle of parsimony justifies picking a simple working hypothesis as a guide to experimenta- tion and assuming that the subject unconsciously went through these steps. But of course further investigation must support the working hypothesis; otherwise, it must eventually be discarded. The divergence of the protocols from the machine trace, as well as the difficulties raised by planning, indicate that things are not so simple as our craving for parsimony leads us to hope. In the light of these difficul- ties, it would be natural to revise the working hypothesis, just as scien- tists had to give up Newtonian Mechanics when it failed to account for certain observations; but at this point, research in Cognitive Simulation deviates from acceptable scientific procedures. In summarizing their work in CS, Newell and Simon conclude: There is a growing body of evidence that the elementary information processes used by the human brain in thinking are highly similar to a subset of the elementary information processes that are incorporated in the instruction codes of the present-day computers. 20 What is this growing body of evidence? Have the gaps in the proto- cols been filled and the exceptions explained? Not at all. The growing body of evidence seems to be the very programs whose lack of univer- sality would cast doubt on the whole project but for the independent assumption of the digital hypothesis. In the face of the exceptions, the psychological assumption would have to already have been taken as independently justified, for the specific programs to be presented as established theories; yet now the assumption is recognized as an hypothe- What Computers Can't Do / 86 sis whose sole confirmation rests on the success of the specific programs. An hypothesis based on a methodological principle is often confirmed later by the facts. What is unusual and inadmissible is that, in this case, the hypothesis produces the evidence by which it is later confirmed. No independent, empirical evidence exists for the psychological as- sumption. In fact, the same empirical evidence presented for the assump- tion that the mind functions like a digital computer tends, when considered without making this assumption, to show that the assumption is empirically untenable. This particular form of methodological confusion is restricted to those working in Cognitive Simulation, but even workers in Artificial Intelli- gence share this belief in the soundness of heuristic programs, this ten- dency to think of all difficulties as accidental, and this refusal to consider any setbacks as discontinuing evidence. Concluding from the small area in which search procedures are partially successful, workers of both schools find it perfectly clear that the unknown and troublesome areas are of exactly the same sort. Thus, all workers proceed as if the credit of the psychological assumption were assured, even if all do not like those in Cognitive Simulation attempt to underwrite the credit with a loan for which it served as collateral. For workers in the field, the psychological assumption seems not to be an empirical hypothesis that can be supported or disconfirmed, but some sort of philosophical axiom whose truth is assured a priori. II. A Priori Arguments for the Psychological Assumption A clue to the a priori character of this axiom can be gained from another look at the way Miller et al. introduce their computer model. The same page which concludes that Simon's success argues strongly for their position opens with a statement of their aims: Any complete description of behavior should be adequate to serve as a set of instructions, that is, it should have the characteristics of a plan that could guide the action described. 21 The Psychological Assumption / 87 Miller et al. assume that our very notion of explanation or complete description requires that behavior be described in terms of a set of instructions, that is, a sequence of determinate responses to determinate situations. No wonder psychologists such as Newell, Neisser, and Miller find work in Cognitive Simulation encouraging. In their view, if psychol- ogy is to be possible at all, an explanation must be expressible as a computer program. This is not an empirical observation but follows from their definition of explanation. Divergences from the protocol and fail- ures can be ignored. No matter how ambiguous the empirical results in Cognitive Simulation, they must be a first step toward a more adequate theory. This definition of explanation clearly needs further investigation: Does it make sense? Even if it does, can one prejudge the results in psychology by insisting theories must be computer programs because otherwise psy- chology isn't possible? Perhaps, psychology as understood by the cogni- tive simulationists is a dead end. To begin with it is by no means clear what the pronouncement that a complete description must take the form of a set of instructions means. Consider the behavior involved in selecting, on command, a red square from a multicolored array of geometrical figures. A complete description of that behavior according to Miller et al. would be a set of instructions, a plan to follow in carrying out this task. What instructions could one give a person about to undertake this action? Perhaps some very general rules such as listen to the instructions, look toward the objects, consider the shapes, make your selection. But what about the detailed instructions for identifying a square rather than a circle? One might say: "Count the sides; if there are four, it is a square." And what about the instructions for identifying a side? "Take random points and see if they fall on a line which is the shortest distance between the end points," and so on. And how does one find these points? After all, there are no points in the field of experience when I am confronting a display of geometrical figures. Perhaps here the instructions run out and one just says: "But you uncon- sciously see points and unconsciously count." But do you? And why do the instructions stop here and not earlier or later? And if all this does What Computers Can't Do / 88 not seem strange enough, what instructions do you give someone for distinguishing red from blue? At this point it is no longer clear why or how a complete description in psychology should take the form of a set of instructions. Still such a claim is the heir to a venerable tradition. Kant explicitly analyzed all experience, even perception, in terms of rules, and the no- tion that knowledge involves a set of explicit instructions is even older. In fact, we have seen that the conviction that a complete description involving an analysis into instructions must be possible, because only such an analysis enables us to understand what is going on, goes back to the beginning of philosophy, that is, to the time when our concepts of understanding and reason were first formulated. Plato, who formu- lated this analysis of understanding in the Euthyphro, goes on to ask in the Meno whether the rules required to make behavior intelligible to the philosopher are necessarily followed by the person who exhibits the behavior. That is, are the rules only necessary if the philosopher is to understand what is going on, or are these rules necessarily followed by the person insofar as he is able to behave intelligently? Since Plato generally thought of most skills as just pragmatic puttering, he no doubt held that rules were not involved in understanding (or producing) skilled behavior. But in the case of theorem proving or of moral action, Plato thought that although people acted without necessarily being aware of any rules, their action did have a rational structure which could be explicated by the philosopher, and he asks whether the mathematician and the moral agent are implicitly following this program when behaving intelligently. This is a decisive issue for the history of our concepts of understanding and explanation. Plato leaves no doubt about his view: any action which is in fact sensible, i.e., nonarbitrary, has a rational structure which can be expressed in terms of some theory and any person taking such action will be following, at least implicitly, this very theory taken as a set of rules. For Plato, these instructions are already in the mind, prepro- grammed in a previous life, and can be made explicit by asking the subjects the appropriate questions. 22 Thus, for Plato, a theory of human behavior which allows us to understand what a certain segment of that The Psychological Assumption / 89 behavior accomplishes is also an explanation of how that behavior is produced. Given this notion of understanding and this identification of understanding and explanation, one is bound to arrive at the cognitive simulationists with their assumption that it is self-evident that a complete description of behavior is a precise set of instructions for a digital com- puter, and that these rules can actually be used to program computers to produce the behavior in question. We have already traced the history of this assumption that thinking is calculating. 23 We have seen that its attraction harks back to the Pla- tonic realization that moral life would be more bearable and knowledge more definitive if it were true. Its plausibility, however, rests only on a confusion between the mechanistic assumptions underlying the success of modern physical science and a correlative formalistic assumption underlying what would be a science of human behavior if such existed. On one level, this a priori assumption makes sense. Man is an object. The success of modern physical science has assured us that a complete description of the behavior of a physical object can be expressed in precise laws, which in turn can serve as instructions to a computer which can then, at least in principle, simulate this behavior. This leads to the idea of a neurophysiological description of human behavior in terms of inputs of energy, physical-chemical transactions in the brain, and outputs in terms of motions of the physical body, all, in principle, simulatable on a digital machine. This level of description makes sense, at least at first approximation, and since the time of Descartes has been part of the idea of a total physical description of all the objects in the universe. The brain is clearly an energy-transforming organ. It detects incoming signals; for example, it detects changes in light intensity correlated with changes in texture gradient. Unfortunately for psychologists, however, this physical de- scription, excluding as it does all psychological terms, is in no way a psychological explanation. On this level one would not be justified in speaking of human agents, the mind, intentions, perceptions, memories, or even of colors or sounds, as psychologists want to do. Energy is being received and transformed and that is the whole story. There is, of course, another level let us call it phenomenological What Computers Can't Do / 90 on which it does make sense to talk of human agents, acting, perceiving objects, and so forth. On this level what one sees are tables, chairs, and other people, what one hears are sounds and sometimes words and sentences, and what one performs are meaningful actions in a context already charged with meaning. But this level of description is no more satisfactory to a psychologist than the physiological level, since here there is no awareness of following instructions or rules; there is no place for a psychological explanation of the sort the cognitive simulationist demands. Faced with this conceptual squeeze, psychologists have always tried to find a third level on which they can do their work, a level which is psychological and yet offers an explanation of behavior. If psychology is to be a science of human behavior, it must study man as an object. But not as a physical object, moving in response to inputs of physical energy, since that is the task of physics and neurophysiology. The alternative is to try to study human behavior as the response of some other sort of object to some other sort of input. Just what this other sort of object and input are is never made clear, but whatever they are, if there is to be an explanation, man must be treated as some device responding to discrete elements, according to laws. These laws can be modeled on causal laws describing how fixed propensities in the organism interact with inputs from the environment to produce complex forms of behavior. The device, then, is a reflex machine, and the laws are the laws of association. This gives us the empiricist psychology of David Hume with its modern descendant, S-R psychology. Or the object can be treated as an information-processing device and the laws can be understood on the Kantian model, as reasons, which are rules in the mind applied by the mind to the input. In psychology this school was called idealist, intellec- tualist, or mentalist, and is now called "cognitive psychology." Until the advent of the computer the empiricist school had the edge because the intellectualist view never succeeded in treating man as a calculable object. There was always a subject, a "transcendental ego," applying the rules, which simply postponed a scientific theory of behav- ior by installing a little man (homunculus) in the mind to guide its actions. Computers, however, offer the irresistible attraction of operating according to rules without appeal to a transcendental ego or homun- The Psychological Assumption / 91 culus. Moreover, computer programs provide a model for the analysis of behavior such as speaking a natural language which seems to be too complex to be accounted for in terms of S-R psychology. In short, there is now a device which can serve as a model for the mentalist view, and it is inevitable that regardless of the validity of the arguments or the persuasiveness of the empirical evidence, psychologists dissatisfied with behaviorism will clutch at this high-powered straw. A computer is a physical object, but to describe its operation, one does not describe the vibrations of the electrons in its transistors, but rather the levels of organization of its on/off flip/flops. If psychological con- cepts can be given an interpretation in terms of the higher levels of organization of these rule-governed flip/flops, then psychology will have found a language in which to explain human behavior. The rewards are so tempting that the basic question, whether this third level between physics and phenomenology is a coherent level of discourse or not, is not even posed. But there are signs of trouble. The language of books such as those by Miller et al., Neisser, and Fodor is literally incoherent. On almost every page one finds sentences such as the fol- lowing: When an organism executes a Plan he proceeds step by step, completing one part and then moving to the next. 24 * Here all three levels exist in unstable and ungrammatical suspension. "When an organism [biological] executes [machine analogy borrowed from human agent] a Plan he [the human agent] . . ." Or, one can have it the other way around and instead of the organism being personified, one can find the mind mechanized. Fodor speaks of "mental process- ing," 25 or "mental operations," 26 as if it were clear what such a form of words could possibly mean. This new form of gibberish would merely be bizarre if it did not reveal more serious underlying conceptual confusions. These are implicit in the work of Miller et al. but become clear in the works of Neisser and Fodor, who, of all the writers in this area, make the greatest effort to articulate their philosophical presuppositions. The confusion can best be brought to light by bearing firmly in mind the neurophysiological and phe- What Computers Can't Do / 92 nomenological levels of description and then trying to locate the psycho- logical level somewhere between these two. In trying to make a place for the information-processing level Neisser tells us: There is certainly a real world of trees and people and cars and even books. . . . However, we have no immediate access to the world nor to any of its properties. 27 This is certainly true insofar as man is regarded as a physical object. 28 * As Neisser puts it, ". . . the sensory input is not the page itself; it is pattern of light rays. . . ." 29 So far so good, but then, Neisser goes on to bring the physical and the phenomenological levels together: "Suitably focussed by the lens . . . the rays fall on the sensitive retina, where they can initiate the neural processes that eventually lead to seeing and reading and remembering. " 30 Here, however, things are by no means obvious. There are two senses of "lead to." Light waves falling on the retina eventually lead to physical and chemical processes in the brain, but in this sequential sense, the light rays and neural processes can never even- tually lead to seeing. 31 * Seeing is not a chemical process; thus it is not a final step in a series of such processes. If, on the other hand, "lead to" means "necessary and sufficient condition for," then, either seeing is the whole chain or something totally different from the chain or any link of it. In either case it is no longer clear why Neisser says we have no immediate access to the perceptual world. Once the neural and phenomenological levels have thus been illegiti- mately amalgamated into one series, which stands between the person and the world, a new vocabulary is required. This no-man's-land is described in terms of "sensory input" and its "transformations." As used here, the term "cognition" refers to all the processes by which the sensory input is transformed, reduced, elaborated, stored, recovered, and used. . . . Such terms as sensation, perception, imagery, retention, recall, problem- solving, and thinking, among many others, refer to hypothetical stages or aspects of cognition. 32 Once a "sensory input" which differs from the world we normally see has been introduced, it seems necessary that our perception be "devel- The Psychological Assumption / 93 oped from," or a "transformation of this "stimulus input." 33 * But what this transformation means depends on the totally ambiguous notion of "stimulus input." If the input is energy, then it is only necessary that it be transformed into other energy the processes in the brain are surely physical from beginning to end. Matter-energy can be transformed, re- duced, elaborated, stored, recovered, and used, but it will never be any- thing but matter-energy. If, however, the stimulus is some sort of primitive perception, as Neisser later seems to suggest "a second stimu- lus will have some effect on how the first brief one is perceived" 34 then we have to know more about what this new percept is. Philosophers have ceased to believe in sense data, and if Neisser has some notion of a primitive percept, it cannot be introduced without a great deal of argu- ment and evidence. Phenomenologically we directly perceive physical objects. We are aware of neither sense data nor light rays. If Neisser wants to shift his notion of input from physical to perceptual, it is up to him to explain what sort of perception he has in mind, and what evidence he has that such a percept, which is neither a pattern of light rays nor a perspectival view of a physical object, exists. 35 * "Information" is the concept which is supposed to rescue us from this confusion. Neisser says "Information is what is transformed, and the structured pattern of its transformation is what we want to under- stand." 36 But as long as the notion of "stimulus input" is ambiguous, it remains unclear what information is and how it is supposed to be related to the "stimulus input," be it energy or direct perception. Finally, in a dazzling display of conceptual confusion, these two inter- dependent and ambiguous notions, "stimulus input" and "information," are combined in the "central assertion" of the book: The central assertion is that seeing, hearing, and remembering are all acts of construction, which may make more or less use of stimulus information [sic] depending on circumstances. The constructive processes are assumed to have two stages of which the first is fast, crude, wholistic, and parallel, while the second is deliberate, attentive, detailed, and sequential. 37 The ambiguity of "stimulus information" and the subsequent incoher- ence of the conceptual framework underlying this approach and its What Computers Can't Do / 94 consequences can best be seen by following a specific example. Let us take Neisser's analysis of the perception of a page. If we see moving objects as unified things, it must be because perception results from an integrative process over time. The same process is surely responsible for the construction of visual objects from the successive "snapshots" taken by the moving eye. 38 The question to be asked here is: What are these snapshots? Are they "patterns of energy" or are they momentary pictures of a page? If they are patterns of energy they are in no sense perceived, and are integrated not by the subject (the perceiver) but by the brain as a physical object. On the other hand, on the phenomenological level, we do not have to integrate distinct snapshots of the page at all. The page is steadily seen, and the notion that it is seen as a series of "snapshots" or "inputs" is an abstraction from this continuously presented page. Of course, this steadily seen page is correlated with some "processing," but not the processing of rudimentary perceptual objects, or "snapshots" which could only give rise to the question of how these elementary perceptual objects were themselves "constructed" but the processing of some fluc- tuating pattern of energy bombarding the eye. This conceptual confusion, which results from trying to define a level of discourse between the physiological and the phenomenological, is even more pronounced in Fodor's work, because he tries even harder to be clear on just these points. In discussing the perception of visual and acoustic patterns Fodor notes that "the concept you have of a face, or a tune, or a shape . . . includes a representation of the formal structure of each of these domains and the act of recognition involves the applica- tion of such information to the integration of current sensory inputs." 39 One wonders again what "sensory input" means here. If the "sensory input" is already a face, or a tune, or a shape, then the job is already done. On the other hand, if the "sensory input" is the physical energy reaching the sense organ, then it is impossible to understand what Fodor means by the "application" of a "concept" or of "information" to the integra- tion of such inputs, since what would integrate such physical energy would surely be further energy transformations. The Psychological Assumption / 95 Of course, if we begged the question and assumed that the brain is a digital computer, then sense could be made of the notion that a concept is a formal structure for organizing data. In that case the "sensory input" would be neither a percept nor a pattern of energy, but a series of bits, and the concept would be a set of instructions for relating these bits to other bits already received, and classifying the result. This would amount to an hypothesis that human behavior can be understood on the model of a digital computer. It would require a theory of just what these bits are and would then have to be evaluated on the basis of empirical evidence. But for Fodor, as for Miller et al., the notion of "sensory input" and of a concept as a rule for organizing this input seems to need no justifica- tion but rather to be contained in the very notion of a psychological explanation. Insofar as it seeks to account for behavior, a psychological theory may be thought of as a function that maps an infinite set of possible inputs to an organism onto an infinite set of possible outputs. 4 ., 40 As a conceptual analysis of the relation of perception and behavior, which is supposed to be accepted independently of empirical assump- tions about the brain, such an account is incomprehensible. As with Neisser, this incoherence can best be seen in a specific case. Fodor takes up the problem of how "we have learned to hear as similar" as one melody "what may be physically quite different sequences of tones." 41 Here the question-begging nature of the analysis is clear: Are these sequences of tones physical or phenomenal? Are they patterns of sound waves or percepts? The talk of their physical difference suggests the former. And indeed on the level of physical energy it is no doubt true that inputs of energy of various frequencies are correlated with the same perceptual experience. The energy transformations involved will pre- sumably someday be discovered by neurophysiologists. But such physical sequences of tones cannot be "heard" we do not hear frequencies; we hear sounds : and thus a fortiori these frequencies cannot be "heard as similar." If, on the other hand, we try to understand the input as se- quences of phenomenal tones, which it would make sense to "hear as What Computers Can't Do / 96 similar," then we are on the level of perception, and unfortunately for Fodor the problem of how we hear these sequences of tones as similar vanishes; for in order to pose the problem in the first place we have already assumed that the phenomenal tone sequences are heard as simi- lar. On the phenomenal level we hear them as similar because they sound similar. To put it another way, Fodor speaks of "which note in particular (i.e., which absolute values of key, duration, intensity, stress, pitch, ampli- tude, etc.) we expect after hearing the first few notes of a performance of Lilliburlero. . . ." 42 But we do not "expect" any "absolute values" at all. We expect notes in a melody. The absolute values pose a problem for the neurophysiologist with his oscilloscope, or for someone hearing the notes in isolation, not for the perceiver. If we did perceive and expect these "absolute values," we would indeed need the "elaborate conceptualism" defended by Fodor, in order to recognize the same melody in various sequences: It is unclear how to account for the ability to recognize identity of type despite gross variations among tokens unless we assume that the concepts employed in recognition are of formidable abstractness. But then it is unclear how the applica- tion of such concepts ... is to be explained, unless one assumes psychological mechanisms whose operations must be complicated in the extreme. 43 Here the confusion shows up in the use of "token" and "type." What are these tokens? The perceived phenomenal sound sequence (the mel- ody) cannot be an abstraction (a type) of which the physical energy in- puts are instantiations (tokens). The percept and the physical energy are equally concrete and are totally different sorts of phenomena. No amount of complication can bridge the gap between shifting energy inputs and the perception of an enduring sound. One is not an instantia- tion of the other. But neither can the tokens be taken to be the phenome- nal sequence of isolated absolute tones (as a sense data theorist would have it). In listening to a melody absolute tones are not perceived, so under this interpretation there would be no tokens at all. Even if one assumes that Fodor has in mind the physical model, which could be computerized, this kind of pattern recognition could conceiva- The Psychological Assumption / 97 bly be accomplished by a neural net or by an analogue device, if it could be accomplished at all. There is no reason to suppose that it is accom- plished by a heuristic program (a set of abstract concepts), let alone that such a program is a conceptual necessity. Yet Fodor never questions the assumption that there is an informa- tion-processing level on which energy transformation can be discussed in terms of a sequence of specific operations. His only question is: How can we tell that we and the machine have the same program, that is, perform the same operations? Thus, for example, after asking how one could know whether one had a successful machine simulation, Fodor replies: ". . . we need only accept the convention that we individuate forms of behavior by reference not solely to the observable gestures output by an organism but also to the sequence of mental operations that underlie those gestures. "** Or even more baldly: strong equivalence requires that the operations that underlie the production of machine behavior be of the same type as the operations that underlie the produc- tion of organic behavior. 45 It should now be clear that Fodor's argument depends on two sorts of assumptions: First, like Miller et al. and Neisser, he introduces the ambiguous notion of "input" to allow a level of description on which it seems to make sense to analyze perception as if man were a computer receiving some sort of data called "stimulus information." This amounts to the assumption that besides energy processing, "data processing is involved in perception." 46 Fodor then makes two further assumptions of a second sort, of which he seems to be unaware: (1) that this data processing takes place as if on a digital computer, that is, consists of discrete operations, and (2) that this digital computer operates serially according to something like a heuristic program, so that one can speak of a sequence of such opera- tions. Fodor's defense of his "elaborate conceptualism," of his notion that perception requires complicated mental operations, seems to turn on thus dogmatically introducing information processing and then simply overlooking all alternative forms of computers and even alternative What Computers Can't Do / 98 forms of digital data processing. This blindness to alternatives can be seen in the conclusion of Fodor's discussion of such phenomena as the recognition of melodies: Characteristically such phenomena have to do with "constancies" that is, cases in which normal perception involves radical and uniform departure from the informational content of the physical input. It has been recognized since Helm- holtz that such cases provide the best argument for unconscious mental opera- tions for there appears to be no alternative to invoking such operations if we are to explain the disparity between input and percept. 41 Fodor's whole discussion of the logic of computer simulation is vi- tiated by his unquestioned reliance on these questionable assumptions. The ease with which his nonarguments pass for conceptual analysis reveals the grip of the Platonic tradition, and the need to believe in the information-processing level if psychology is to be a science. Of course, the use of the computer as a model is legitimate as long as it is recognized as an hypothesis. But in the writing of Miller et al., Neisser, and Fodor, as we have seen, this hypothesis is treated as an a priori truth, as if it were the result of a conceptual analysis of behavior. Occasionally one glimpses an empirical basis for this assumption: Fodor's argument for the legitimacy of a computer program as a psycho- logical theory ultimately rests on the hypothetical supposition "that we have a machine that satisfies whatever experimental tests we can devise for correspondences between its repertoire and that of some organism." 48 However, this covertly empirical character of the argument is implicitly denied since the whole discussion is couched in terms of "sequences of mental operations," as if it were already certain that such a machine could exist. Only if such a machine existed, and only if it did indeed operate in sequences of steps, would one be justified in using the notions connected with heuristically programmed digital computers to suggest and inter- pret experiments in psychology. But to decide whether such an intelligent machine can exist, and therefore whether such a conceptual framework is legitimate, one must first try to program such a machine, or evaluate the programs already tried. To use computer language as a self-evident The Psychological Assumption / 99 and unquestionable way of formulating the conceptional framework in terms of which experiments are undertaken and understood without valid a priori arguments or an empirical existence-proof of the possibility of such a machine, can only lead to confusion. Conclusion So we again find ourselves moving in a vicious circle. We saw at the end of Section I of this chapter that the empirical results, riddled with unexplained exceptions, and unable to simulate higher-order processes such as zeroing in and essential/inessential discrimination, are only promising if viewed in terms of an a priori assumption that the mind must work like a heuristically programmed digital computer. But now we have seen that the only legitimate argument for the assumption that the mind functions like a computer turns on the actual or possible existence of such an intelligent machine. The answer to the question whether man can make such a machine must rest on the evidence of work being done. And on the basis of actual achievements and current stagnation, the most plausible answer seems to be, No. It is impossible to process an indifferent "input" without distinguishing between relevant and irrelevant, significant and insignifi- cant data. We have seen how Newell, Shaw, and Simon have been able to avoid this problem only by predigesting the data, and how Miller et al. have been able to avoid it only by mistakenly supposing that Newell, Shaw, and Simon had a program which performed this original selection. But if there is no promising empirical evidence, the whole self-supporting argument tumbles down like a house of cards. The only alternative way to cope with selectivity would be analogue processing, corresponding to the selectivity of our sense organs. But then all processing would no longer be digital, and one would have reason to wonder whether this analogue processing was only peripheral. All of which would cast doubt on the "sequence of operations" and reopen the whole discussion. These difficulties suggest that, although man is surely a physical object processing physical inputs according to the laws of What Computers Can't Do / 700 physics and chemistry, man's behavior may not be explainable in terms of an information-processing mechanism receiving and processing a set of discrete inputs. Moreover, nothing from physics or experience sug- gests that man's actions can be so explained, since on the physical level we are confronted with continuously changing patterns of energy, and on the phenomenological level with objects in an already organized field of experience. An analysis of this field of experience would provide an alternative area of study for psychology. But before we turn to this alternative the- ory in Part III, we must follow up two other assumptions, which, even if work in CS cannot be defended, seem to lend plausibility to work in AI. The Epistemological Assumption It should now be evident that it is extremely difficult to define what the mental level of functioning is, and that whatever the mind is, it is by no means obvious that it functions like a digital computer. This makes practically unintelligible the claims of those working in Cognitive Simu- lation that the mind can be understood as processing information accord- ing to heuristic rules. The computer model turns out not to be helpful in explaining what people actually do when they think and perceive, and, conversely, the fact that people do think and perceive can provide no grounds for optimism for those trying to reproduce human performance with digital computers. But this still leaves open another ground for optimism: although hu- man performance might not be explainable by supposing that people are actually following heuristic rules in a sequence of unconscious opera- tions, intelligent behavior may still be formalizable in terms of such rules and thus reproduced by machine. 1 * This is the epistemological assump- tion. Consider the planets. They are not solving differential equations as they swing around the sun. They are not following any rules at all; but their behavior is nonetheless lawful, and to understand their behavior we find a formalism in this case differential equations which expresses their behavior as motion according to a rule. Or, to take another example: What Computers Can't Do / 102 A man riding a bicycle may be keeping his balance just by shifting his weight to compensate for his tendency to fall. The intelligible content of what he is doing, however, might be expressed according to the rule: wind along a series of curves, the curvature of which is inversely propor- tional to the square of the velocity. 2 * The bicycle rider is certainly not following this rule consciously, and there is no reason to suppose he is following it unconsciously. Yet this formalization enables us to express or understand his competence, that is, what he can accomplish. It is, however, in no way an explanation of his performance. It tells us what it is to ride a bicycle successfully, but nothing of what is going on in his brain or in his mind when he performs the task. There is thus a subtle but important difference between the psychologi- cal and the epistemological assumptions. Both assume the Platonic no- tion of understanding as formalization, but those who make the psychological assumption (those in CS) suppose that the rules used in the formalization of behavior are the very same rules which produce the behavior, while those who make the epistemological assumption (those in AI) only affirm that all nonarbitrary behavior can be formalized according to some rules, and that these rules, whatever they are, can then be used by a computer to reproduce the behavior. The epistemological assumption is weaker and thus less vulnerable than the psychological assumption. But it is vulnerable nonetheless. Those who fall back on the epistemological assumption have realized that their formalism, as a theory of competence, need not be a theory of human performance, but they have not freed themselves sufficiently from Plato to see that a theory of competence may not be adequate as a theory of machine performance either. Thus, the epistemological as- sumption involves two claims: (a) that all nonarbitrary behavior can be formalized, and (b) that the formalism can be used to reproduce the behavior in question. In this chapter we shall criticize claim (a) by showing that it is an unjustified generalization from physical science, and claim (b) by trying to show that a theory of competence cannot be a theory of performance: that unlike the technological application of the laws of physics to produce physical phenomena, a timeless, contextless theory of competence cannot be used to reproduce the moment-to- The Epistemological Assumption / 103 moment involved behavior required for human performance; that indeed there cannot be a theory of human performance. If this argument is convincing, the epistemological assumption, in the form in which it seems to support AI, turns out to be untenable, and, correctly under- stood, argues against the possibility of AI, rather than guaranteeing its success. Claim (a), that all nonarbitrary behavior can be formalized, is not an axiom. It rather expresses a certain conception of understanding which is deeply rooted in our culture but may nonetheless turn out to be mistaken. We must now turn to the empirical arguments which can be given in support of such a hypothesis. It should also be clear by now that no empirical arguments from the success of AI are acceptable, since it is precisely the interpretation, and, above all, the possibility of significant extension of the meager results such as Bobrow's which is in question. Since two areas of successful formalization physics and linguistics seem to support the epistemological assumption, we shall have to study both these areas. In physics we indeed find a formalism which describes behavior (for example, the planets circling the sun), but we shall see that this sort of formalism can be of no help to those working in AI. In linguistics we shall find, on the other hand, a formalism which is relevant to work in AI, and which argues for the assumption that all nonarbitrary behavior can be formalized, but we will find that this formalism which expresses the competence of the speaker that is, what he is able to accomplish cannot enable one to use a computer to reproduce his performance that is, his accomplishment. I. A Mistaken Argument from the Success of Physics Minsky's optimism that is, his conviction that all nonarbitrary behav- ior can be formalized and the resulting formalism used by a digital computer to reproduce that behavior is a pure case of the epistemologi- cal assumption. It is this belief which allows Minsky to assert with confidence that "there is no reason to suppose that machines have any limitations not shared by man." 3 We must now examine the arguments What Computers Can't Do / 104 supporting this claim, but first we must be clear what the formalist means by machine. A digital computer is a machine which operates according to the sort of criteria Plato once assumed could be used to understand any orderly behavior. This machine, as defined by Minsky, who bases his definition on that of Turing, is a "rule-obeying mechanism." As Turing puts it: "The . . . computer is supposed to be following fixed rules. ... It is the duty of the control to see that these instructions are obeyed correctly and in the right order. The control is so constructed that this necessarily happens." 4 So the machine in question is a restricted but very fundamen- tal sort of mechanism. It operates on determinate, unambiguous bits of data, according to strict rules which apply unequivocally to these data. The claim is made that this sort of machine a Turing machine which expresses the essence of a digital computer can, in principle, do anything that human beings can do that it has, in principle, only those limita- tions shared by man. Minsky considers the antiformalist counterclaim that "perhaps there are processes . . . which simply cannot be described in any formal language, but which can nevertheless be carried out, e.g., by minds." 5 Rather than answer this objection directly, he refers to Turing's "bril- liant" article which, he asserts, contains arguments that "amount ... to a satisfactory refutation of many such objections." 6 Turing does, indeed, take up this sort of objection. He states it as follows: "It is not possible to produce a set of rules purporting to describe what a man should do in every conceivable set of circumstances." 7 This is presuma- bly Turing's generalization of Wittgenstein's argument that it is impossi- ble to supply normative rules which prescribe in advance the correct use of a word in all situations. Turing's "refutation" is to make a distinction between "rules of conduct" and "laws of behavior" and then to assert that "we cannot so easily convince ourselves of the absence of complete laws of behavior as of complete rules of conduct." 8 Now as an answer to the Wittgensteinian claim, this is well taken. Turing is in effect arguing that although we cannot formulate the norma- tive rules for the correct application of a particular predicate, this does The Epistemological Assumption / 705 not show that we cannot formulate the rules which describe how, in fact. a particular individual applies such a predicate. In other words, while Turing is ready to admit that it may in principle be impossible to provide a set of rules describing what a person should do in every circumstance, he holds there is no reason to doubt that one could in principle discover a set of rules describing what he would do. But why does this supposition seem so self-evident that the burden of proof is on those who call it into question? Why should we have to "convince ourselves of the absence of complete laws of behavior" rather than of their presence? Here we are face to face again with the epistemological assumption. It is important to try to root out what lends this assumption its implied a priori plausi- bility. To begin with, "laws of behavior" is ambiguous. In one sense human behavior is certainly lawful, if lawful simply means orderly. But the assumption that the laws in question are the sort that could be embodied in a computer program or some equivalent formalism is a different and much stronger claim, in need of further justification. The idea that any description of behavior can be formalized in a way appropriate to computer programming leads workers in the field of artificial intelligence to overlook this question. It is assumed that, in principle at least, human behavior can be represented by a set of indepen- dent propositions describing the inputs to the organism, correlated with a set of propositions describing its outputs. The clearest statement of this assumption can be found in James Culbertson's move from the assertion that in theory at least one could build a robot using only flip/flops to the claim that it could therefore reproduce all human behavior. Using suitable receptors and effectors we can connect them together via central cells. If we could get enough central cells and if they were small enough and if each cell had enough endbulbs and if we could put enough bulbs at each synapse and if we had time enough to assemble them, then we could construct robots to satisfy any given input-output specification, i.e., we could construct robots that would behave in any way we desired under any environmental circumstances. There would be no difficulty in constructing a robot with behavioral properties just like John Jones or Henry Smith or in constructing a robot with any desired behavioral improvements over Jones and Smith. 9 What Computers Can't Do / 106 Or put more baldly: Since [these complete robots] can, in principle, satisfy any given input-output specifications, they can do any prescribed things under any prescribed circum- stances ingeniously solve problems, compose symphonies, create works of art and literature and engineering, and pursue any goals. 10 But as we have seen in Chapter 4, it is not clear in the case of human beings what these inputs and outputs are supposed to be. 11 * Culbertson's assumption that the brain can be understood as correlating isolated bits of data rests on the assumption that the neurons act as on/off switches. Since, as we have seen in Chapter 3, this is probably not the case, there is no reason to suppose, and several reasons to doubt, that human inputs and outputs can be isolated and their correlation formalized. Culbert- son's assumption is an assumption and nothing more, and so in no way justifies his conclusions. The committed formalist, however, has one more move. He can ex- ploit the ambiguity of the notion of "laws of behavior," and take behavior to mean not meaningful human actions, but simply the physical move- ments of the human organism. Then, since human bodies are part of the physical world and, as we have seen, objects in the physical world have been shown to obey laws which can be expressed in a formalism manipu- lable on a digital computer, the formalist can still claim that there must be laws of human behavior of the sort required by his formalism. To be more specific, if the nervous system obeys the laws of physics and chemis- try, which we have every reason to suppose it does, then even if it is not a digital computer, and even if there is no input-output function directly describing the behavior of the human being, we still ought to be able to reproduce the behavior of the nervous system with some physical device which might, for example, take the form of a new sort of "analogue computer" using ion solutions whose electrical properties change with various local saturations. Then, as we pointed out in Chapter 4, knowing the composition of the solutions in this device would enable us at least in principle to write the physicochemical equations describing such wet components and to solve these equations on a dry digital computer. The Epistemological Assumption / 707 Thus, given enough memory and time, any computer even such a special sort of analogue computer could be simulated on a digital ma- chine. In general, by accepting the fundamental assumptions that the nervous system is part of the physical world and that all physical pro- cesses can be described in a mathematical formalism which can in turn be manipulated by a digital computer, one can arrive at the strong claim that the behavior which results from human "information processing," whether directly formalizable or not, can always be indirectly repro- duced on a digital machine. This claim may well account for the formalist's smugness, but what in fact is justified by the fundamental truth that every form of "informa- tion processing" (even those which in practice can only be carried out on an "analogue computer") must in principle be simulable on a digital computer? We have seen it does not prove the mentalist claim that, even when a human being is unaware of using discrete operations in process- ing information, he must nonetheless be unconsciously following a set of instructions. Does it justify the epistemological assumption that all nonarbitrary behavior can be formalized? One must delimit what can count as information processing in a computer. A digital computer solving the equations describing an ana- logue information-processing device and thus simulating its function is not thereby simulating its "information processing." It is not processing the information which is processed by the simulated analogue, but en- tirely different information concerning the physical or chemical proper- ties of the analogue. Thus the strong claim that every form of information can be processed by a digital computer is misleading. One can only show that for any given type of information a digital computer can in principle be programmed to simulate a device which can process that information. Thus understood as motion as the input and output of physical signals human behavior is presumably completely lawful in the sense the formalists require. But this in no way supports the formalist assump- tion as it appears in Minsky and Turing. For when Minsky and Turing claim that man is a Turing machine, they cannot mean that a man is a physical system. Otherwise it would be appropriate to say that planes or What Computers Can't Do / 108 boats are Turing machines. Their behavior, too, can be described by mathematically formulable laws relating their intake and output of energy and can at least in principle be reproduced to any degree of accuracy on a digital computer. No, when Minsky or Turing claims that man can be understood as a Turing machine, they must mean that a digital computer can reproduce human behavior, not by solving physical equations, but by processing data received from the world, by means of logical operations that can be reduced to matching, classifying, and Boolean operations. As Minsky puts it: Mental processes resemble ... the kinds of processes found in computer pro- grams: arbitrary symbol associations, treelike storage schemes, conditional trans- fers, and the like. 12 Workers in AI are claiming that this mental level of "information pro- cessing" can be described in a digital formalism. All AI research is dedicated to using logical operations to manipulate data directly from the world, not to solving physical equations describing physical objects. Considerations from physics show only that inputs of energy, and the neurological activity involved in transforming them, can in principle be described in this digital form. No one has tried, or hopes to try, to use the laws of physics to calculate in detail the motion of human bodies. Indeed, this may well be physically impossible, for H. J. Bremermann has shown that: No data processing system whether artificial or living can process more than (2 X 10 47 ) bits per second per gram of its mass. 13 Bremermann goes on to draw the following conclusions: There are\7T X 10 7 seconds in a year. The age of the earth is about 10 9 years, its mass less than 6 X 10 27 grams. Hence even a computer of the size of the earth could not process more than 10" bits during a time equal to the age of the earth. [Not to mention the fact, one might add, that the bigger the computer the more the speed of light would be a factor in slowing down its operation.] . . . Theorem proving and problem solving . . . lead to exponentially growing problem trees. If our conjecture is true then it seems that the difficulties that are currently encountered in the field of pattern recognition and theorem proving will not be resolved by sheer speed of data processing by some future super-computers. 14 The Epistemological Assumption / 709 If these calculations are correct, there is a special kind of impossibility involved in any attempt to simulate the brain as a physical system. The enormous calculations necessary may be precluded by the very laws of physics and information theory such calculations presuppose. Yet workers in the field of AI from Turing to Minsky seem to take refuge in this confusion between physical laws and information-process- ing rules to convince themselves that there is reason to suppose that human behavior can be formalized; that the burden of proof is on those who claim that "there are processes . . . which simply cannot be described in a formal language but which can nevertheless be carried out, e.g., by minds." 15 Once we have set straight the equivocation between physical laws and information-processing rules, what argument remains that hu- man behavior, at what AI workers have called "the information process- ing level," can be described in terms of strict rules? II. A Mistaken Argument from the Success of Modern Linguistics If no argument based on the success of physics is relevant to the success of AI, because AI is concerned with formalizing human behavior not physical motion, the only hope is to turn to areas of the behavioral sciences themselves. Galileo was able to found modern physics by ab- stracting from many of the properties and relations of Aristotelian phys- ics and finding that the mathematical relations which remained were sufficient to describe the motion of objects. What would be needed to justify the formalists' optimism would be a Galileo of the mind who, by making the right abstractions, could find a formalism which would be sufficient to describe human behavior. John McCarthy expresses this longing for a rapprochement between physics and the behavioral sciences: Although formalized theories have been devised to express the most important fields of mathematics and some progress has been made in formalizing certain empirical sciences, there is at present no formal theory in which one can express the kind of means-ends analysis used in ordinary life. . . . Our approach to the artificial-intelligence problem requires a formal theory. 16 What Computers Can't Do / 110 Recently such a breakthrough has occurred. Chomsky and the trans- formational linguists have found that by abstracting from human perfor- mance the use of particular sentences on particular occasions they can formalize what remains, that is, the human ability to recognize gram- matically well-formed sentences and to reject ill-formed ones. That is, they can provide a formal theory of much of linguistic competence. 11 * This success is a major source of encouragement for those in AI who are committed to the view that human behavior can be formalized without reduction to the physical level. For such success tends to confirm at least the first half of the epistemological hypothesis. A segment of orderly behavior which at first seems nonrulelike turns out to be describable in terms of complex rules, rules of the sort which can be processed directly by a digital computer (directly that is, without passing by way of a physical description of the motions of the vocal cords of a speaker or the physiochemical processes taking place in his brain). But such a formalization only provides justification for half the epis- temological hypothesis. Linguistic competence is not what AI workers wish to formalize. If machines are to communicate in natural language, their programs must not only incorporate the rules of grammar; they must also contain rules of linguistic performance. In other words, what was omitted in order to be able to formulate linguistic theory the fact that people are able to use their language is just what must also be formalized. The question whether the epistemological hypothesis is justified thus comes down to the test case: is there reason to suppose that there can be a formal theory of linguistic performance? There are two sorts of reasons to believe that such a generalization of linguistic theory is impos- sible: (1) An argument of principle (to which we shall turn in the next chapter): for there to be a theory of linguistic performance, one would have to have a theory of all human knowledge; but this, it seems, is impossible. (2) A descriptive objection (to which we shall now turn): not all linguistic behavior is rulelike. We recognize some linguistic expres- sions as odd as breaking the rules and yet we are able to understand them. There are cases in which a native speaker recognizes that a certain The Epistemological Assumption / 777 linguistic usage is odd and yet is able to understand it for example, the phrase "The idea is in the pen" is clear in a situation in which we are discussing promising authors; but a machine at this point, with rules for what size physical objects can be in pig pens, playpens, and fountain pens, would not be able to go on. Since an idea is not a physical object, the machine could only deny that it could be in the pen or at best make an arbitrary stab at interpretation. The listener's understanding, on the other hand, is far from arbitrary. Knowing what he does about the shadow which often falls between human projects and their execution, as well as what one uses to write books, he gets the point, and the speaker will often agree on the basis of the listener's response that the listener has understood. Does it follow, then, that in understanding or using the odd utterance, the human speakers were acting according to a rule in this case a rule for how to modify the meaning of "in"? It certainly does not seem so to the speakers who have just recognized the utterance as "odd." This case takes us to the heart of a fundamental difficulty facing the simulators. Programmed behavior is either arbitrary or strictly rulelike. Therefore, in confronting a new usage a machine must either treat it as a clear case falling under the rules, or take a blind stab. A native speaker feels he has a third alternative. He can recognize the usage as odd, not falling under the rules, and yet he can make sense of it give it a meaning in the context of human life in an apparently nonrulelike and yet nonar- bitrary way. Outright misuse of language demonstrates an even more extreme form of this ability. People often understand each other even when one of the speakers makes a grammatical or semantic mistake. The utterance may not only be outside the rules but actually proscribed by them, and yet such violations often go unnoticed, so easily are they understood. Human beings confronted with these odd cases and outright errors adapt as they go along and then may reflect on the revisions they have made. A machine has either to fail first and then, when given the correct answer, revise its rules to take account of this new usage, or it would have to have all the rules even the rules for how to break the rules and still be understood built into it beforehand. To adopt the first approach, What Computers Can't Do / 772 failing first and revising later, would be to admit that in principle, not just in practice, machines must always trail behind men that they could not be humanly intelligent. To assume, on the other hand, that the rules covering all cases must be explicitly built in or learned since this is the only way a digital computer could simulate the human ability to cope with odd uses runs counter to logic and experience. Logically, it is hard to see how one could formulate the rules for how one could intelligibly break the rules; for, no matter what metarules are formulated, it seems intuitively obvious that the native speaker could break them too and count on the context to get his meaning across to another speaker. Thus no matter what order of metarules one chooses, it seems there will be a higher order of tacit understanding about how to break those rules and still be understood. Phenomenologically, or empirically, the postulation of a set of uncon- scious metarules of which we are not aware leads to other difficulties. Just as in chess the acceptance of the digital model led to the assumption that the chess player must be using unconscious heuristics, even when the player reported that he was zeroing in on patterns of strength and weakness, the assumption of the pre-existence of rules for disambigua- tion introduces a process of which we have no experiential evidence, and fails to take seriously our sense of the oddness of certain uses. And here, as in the case of chess, this flouting of phenomenological evidence leads to a teleological puzzle: Why, if every understandable use of language is covered by rule, should some of these uses appear odd to us? So odd, indeed, that we cannot supply any rule to explain our interpretation. Why, if we have such a battery of rules and lightning-fast capacity for using them on the unconscious level, should we be left consciously perplexed in certain cases and find them peculiar even after we have understood them? These considerations suggest that, although a general theory of syntax and semantic competence can be scientific because it is a timeless for- malism which makes no claim to formalize the understanding of lan- guage in specific situations, contradictions arise when one demands a comparable formalism for linguistic use. These difficulties do not disturb the linguists who, like true scientists, The Epistemological Assumption / 113 carefully limit themselves to linguistic competence, that is, the general principles which apply to all cases, and exclude as extralinguistic our ability to deal with specific situations as they occur. As Kierkegaard points out in his Concluding Unscientific Postscript, the laws of science are universal and timeless, treating all experience as if it could as well be in the past. 18 AI workers, however, want their machines to interact with people in present real-life situations in which objects have special local significance. But computers are not involved in a situation. Every bit of data always has the same value. True, computers are not what Kant would call "transcendentally stupid"; they can apply a rule to a specific case if the specific case is already unambiguously described in terms of general features mentioned in the rule. They can thus simulate one kind of theoretical understanding. But machines lack practical intel- ligence. They are "existentially" stupid in that they cannot cope with specific situations. Thus they cannot accept ambiguity and the breaking of rules until the rules for dealing with the deviations have been so completely specified that the ambiguity has disappeared. To overcome this disability, AI workers must develop an a-temporal, nonlocal, theory of ongoing human activity. The originality, the importance, and the curse of work in machine communication in a natural language is that the machine must use its formalism to cope with real-life situations as they occur. It must deal with phenomena which belong to the situational world of human beings as if these phenomena belonged to the objective formal universe of science. The believer in machine understanding and use of natural language who is encouraged by the success of linguistics is not laboring under a mis- conception about the way consciousness functions, but rather under a misconception about the relation between theoretical and practical understanding. He supposes that one can understand the practical world of an involved active individual in the same terms one can understand the objective universe of science. In short, he claims, as Leibniz first claimed, that one can have a theory of practice. But such an applied theory could not be the same as the technological application of a physical theory, which it seems to parallel. When one uses the laws of physics to guide missiles, for example, the present What Computers Can't Do / 114 performance of the missile is an instantiation of timeless, universal laws which make no reference to the situation except in terms of such laws. But in linguistics, as we have seen, speakers using the language take for granted common situational assumptions and goals. Thus the general laws of competence cannot be directly applied to simulate behavior. To get from the linguistic formalism to specific performance, one has to take into account the speaker's understanding of his situation. If there could be an autonomous theory of performance, it would have to be an entirely new kind of theory, a theory for a local context which described this context entirely in universal yet nonphysical terms. Neither physics nor linguistics offers any precedent for such a theory, nor any comforting assurance that such a theory can be found. Conclusion But to refute the epistemological assumption that there must be a theory of practical activity in the case of language, to deny that the rules governing the use of actual utterances can in principle be completely formalized it is not sufficient to point out that thus far no adequate language translation system has been developed, or that our language is used in flexible and apparently nonrulelike ways. The formalizer can offer the Platonic retort that our failure to formalize our ability to use language shows only that we have not fully understood this behavior; have not yet found the rules for completely describing linguistic perfor- mance. 19 * This defense might at first seem to be similar to the heuristic program- mer's assurance that he will someday find the heuristics which will enable a machine to play chess, even if he has not yet found them. But there is an important difference. The heuristic programmer's confidence is based on an unfounded psychological assumption concerning the way the mind processes information, whereas the formalist's claim is based on a correct understanding of the nature of scientific explanation. To the extent that we have not specified our behavior in terms of unique and precisely defined reactions to precisely defined objects in universally defined situations, we have not understood that behavior in the only The Epistemological Assumption / 115 sense of "understanding" appropriate to science. To answer this a priori claim of the theoretical understanding one cannot counter with a phenomenological description. One must show that the theoretical claim is untenable on its own terms: that the rules which enable a native speaker to speak cannot be completely formalized; that the epistemological assumption is not only implausible but leads to contradictions. Wittgenstein was perhaps the first philosopher since Pascal to note: "In general we don't use language according to strict rules it hasn't been taught us by means of strict rules either." 20 But Wittgenstein did not base his argument against the claim that language was a calculus solely on a phenomenological description of the nonrulelike use of lan- guage. His strongest argument is a dialectical one, based on a regress of rules. He assumes, like the intellectualist philosophers he is criticizing, that all nonarbitrary behavior must be rulelike, and then reduces this assumption to absurdity by asking for the rules which we use in applying the rules, and so forth. Here it is no longer a question of always being able to break the rules and still be understood. After all, we only feel we can go on breaking the rules indefinitely. We might be mistaken. It is a question of whether a complete understanding of behavior in terms of rules is intelligible. Wittgenstein is arguing, as Aristotle argued against Plato, that there must always be a place for interpretation. And this is not, as Turing seemed to think, merely a question of whether there are rules governing what we should do, which can legitimately be ignoreci. It is a question of whether there can be rules even describing what speakers in fact do. To have a complete theory of what speakers are able to do, one must not only have grammatical and semantic rules but further rules which would enable a person or a machine to recognize the context in which the rules must be applied. Thus there must be rules for recognizing the situation, the intentions of the speakers, and so forth. But if the theory then requires further rules in order to explain how these rules are applied, as the pure intellectualist viewpoint would suggest, we are in an infinite regress. Since we do manage to use language, this regress cannot be a What Computers Can't Do / 116 problem for human beings. If AI is to be possible, it must also not be a problem for machines. Both Wittgenstein and the computer theorists must agree that there is some level at which the rules are simply applied and one no longer needs rules to guide their application. Wittgenstein and the AI theorists differ fundamentally, however, on how to describe this stopping point. For Wittgenstein there is no absolute stopping point; we just fill in as many rules as are necessary for the practical demands of the situation. At some level, depending on what we are trying to do, the interpretation of the rule is simply evident and the regress stops. 21 * For the computer people the regress also stops with an interpretation which is self-evident, but this interpretation has nothing to do with the demands of the situation. It cannot, for the computer is not in a situation. It generates no local context. The computer theorist's solution is to build the machine to respond to ultimate bits of context-free, completely deter- minate data which require no further interpretation in order to be under- stood. Once the data are in the machine, all processing must be rulelike, but in reading in the data there is a direct response to determinate features of the machine's environment as, for example, holes in cards or the mosaic of a TV camera, so on this ultimate level the machine does not need rules for applying its rules. Just as the feeding behavior of the baby herring gull is triggered by a red spot and the frog's eye automati- cally signals the presence of a moving black spot, so human behavior, if it is to be completely understood and computerized, must be understood as if triggered by specific features of the environment. As a theory of human psychology (CS) this is surely not a plausible hypothesis. Our sense of oddness of deviant linguistic uses, as well as our feeling that there is nothing in the environment to which we have an inevitable and invariable response, argue against this view. Moreover, as a theory of our "practical competence" (no matter how we actually produce our behavior), this hypothesis is no more attractive. The general adaptability of our language, which enables us to modify meanings and invent analogies, as well as the general flexibility of human and even higher animal behavior, are incomprehensible on this view. Still, these objections are all based on appearances. They are plausible, but not The Epistemological Assumption / 777 necessarily convincing to those committed to the epistemological as- sumption. A full refutation of the epistemological assumption would require an argument that the world cannot be analyzed in terms of determinate data. Then, since the assumption that there are basic unambiguous ele- ments is the only way to save the epistemological assumption from the regress of rules, the formalist, caught between the impossibility of always having rules for the application of rules and the impossibility of finding ultimate unambiguous data, would have to abandon the epistemological assumption altogether. This assumption that the world can be exhaustively analyzed in terms of determinate data or atomic facts is the deepest assumption underlying work in AI and the whole philosophical tradition. We shall call it the ontological assumption, and now turn to analyzing its attraction and its difficulties. The Ontological Assumption Up to now we have been seeking in vain the arguments and evidence that the mind processes information in a sequence of discrete steps like a heuristically programmed digital computer, or that human behavior can be formalized in these terms. We have seen that there are four types of human information processing (fringe consciousness, ambiguity toler- ance, essential/inessential discrimination, and perspicuous grouping), which have resisted formalization in terms of heuristic rules. And we have seen that the biological, psychological, and epistemological assump- tions which allow workers to view these difficulties as temporary are totally unjustified and may well be untenable. Now we turn to an even more fundamental difficulty facing those who hope to use digital comput- ers to produce artificial intelligence: the data with which the computer must operate if it is to perceive, speak, and in general behave intelli- gently, must be discrete, explicit, and determinate; otherwise, it will not be the sort of information which can be given to the computer so as to be processed by rule. Yet there is no reason to suppose that such data about the human world are available to the computer and several reasons to suggest that no such data exist. The ontological assumption that everything essential to intelligent behavior must in principle be understandable in terms of a set of determi- nate independent elements allows AI researchers to overlook this prob- The Ontological Assumption / 779 lem. We shall soon see that this assumption lies at the basis of all thinking in AI, and that it can seem so self-evident that it is never made explicit or questioned. As in the case of the epistemological assumption, we shall see that this conviction concerning the indubitability of what in fact is only an hypothesis reflects two thousand years of philosophical tradition reinforced by a misinterpretation of the success of the physical sciences. Once this hypothesis is made explicit and called into question, it turns out that no arguments have been brought forward in its defense and that, when used as the basis for a theory of practice such as AI, the ontological assumption leads to profound conceptual difficulties. In his introduction to Semantic Information Processing, Minsky warns against the dreadfully misleading set of concepts that people get when they are told (with the best intentions) that computers are nothing but assemblies of flip-flops; that their programs are really nothing but sequences of operations upon binary num- bers, and so on. 1 He tries to combat this discouraging way of looking at digital computers: While this is one useful viewpoint, it is equally correct to say that the computer is nothing but an assembly of symbol-association and process-controlling ele- ments and that programs are nothing but networks of interlocking goal-formulat- ing and means-ends evaluation processes. This latter attitude is actually much healthier because it reduces one's egotistical tendency to assume total compre- hension of all the possible future implications. 2 But Minsky sees only half the difficulty arising from the restriction that the computer must operate on determinate, independent elements. It may indeed be true that one can formulate higher-order rules for the operation of a computer so that the fact that there are flip/flops never appears in the flow chart, that is, on the information-processing level. (On this level, as we have seen in the preceding two chapters, trouble arises because there must always be explicit rules, not because these rules must ultimately be a sequence of operations on binary numbers.) The flip/flops only become a problem when we consider the kind of informa- tion the machine can be given. 3 * We have seen that Newell quite frankly described GPS a program whose information-processing level is cor- What Computers Can't Do / 120 rectly described in terms of interlocking goals and means-ends as "a program for accepting a task environment defined in terms of discrete objects." 4 It is these discrete objects which must feed the flip/flops, or else must be analyzed into further discrete elements in order to do so. Every program for a digital computer must receive its data in this dis- crete form. This raises a special problem, or, more exactly, it creates a problem by determining the way all questions concerning giving information to computers must be raised. Stated in a neutral way the problem is this: as we have seen, in order to understand an utterance, structure a prob- lem, or recognize a pattern, a computer must select and interpret its data in terms of a context. But how are we to impart this context itself to the computer? The sharpest statement of this problem still in neutral terms occurs in Eden's evaluation of work in handwriting recognition: . . . when [a human being] reads a letter written in a difficult script ... he can reconstruct it with the help of his knowledge of the grammar of the language, the meaning of the text he has been able to read, the character of the subject matter, and, perhaps, the state of mind of the writer. There is now, alas, no hint of how to embody such knowledge of the world and its ways in the computer. 5 Here Eden wisely takes no stand on what we know when we have "knowledge of the world and its ways." The flip/flops secretly come in, however, and, along with the ontological assumption, dictate an answer to this question which is no longer neutral, but rather embodies the computer's requirements. When one asks what this knowledge of the world is, the answer comes back that it must be a great mass of discrete facts. Thus at the end of his introduction to Semantic Information Process- ing, when Minsky finally asks "what is the magnitude of the mass of knowledge required for a humanoid intelligence?" 6 he has already pre- judged the question and unhesitatingly answers in terms of numbers of facts: If we discount specialized knowledge and ask instead about the common-every- day structures that which a person needs to have ordinary common sense we will find first a collection of indispensable categories, each rather complex: geo- The Ontological Assumption / 727 metrical and mechanical properties of things and of space; uses and properties of a few thousand objects; hundreds of "facts" about hundreds of people, thou- sands of facts about tens of people, tens of facts about thousands of people; hundreds of facts about hundreds of organizations. As one tries to classify all his knowledge, the categories grow rapidly at first, but after a while one encounters more and more difficulty. My impression, for what it's worth, is that one can find fewer than ten areas each with more than ten thousand "links." One can't find a hundred things that he knows a thousand things about. Or a thousand things each with a full hundred new links. I therefore feel that a machine will quite critically need to acquire the order of a hundred thousand elements of knowledge in order to behave with reasonable sensibility in ordinary situations. A million, if properly organized, should be enough for a very great intelligence. If my argument does not convince you, multiply the figures by ten. 7 Granting for the moment that all human knowledge can be analyzed as a list of objects and of facts about each, Minsky's analysis raises the problem of how such a large mass of facts is to be stored and accessed, How could one structure these data a hundred thousand discrete ele- ments so that one could find the information required in a reasonable amount of time? When one assumes that our knowledge of the world is knowledge of millions of discrete facts, the problem of artificial intelli- gence becomes the problem of storing and accessing a large data base. Minsky sees that this presents grave difficulties: ... As everyone knows, it is hard to find a knowledge-classifying system that works well for many different kinds of problems: it requires immense effort to build a plausible thesaurus that works even within one field. Furthermore, any particular retrieval structure is liable to entail commitments making it difficult to incorporate concepts that appear after the original structure is assembled. One is tempted to say: "It would be folly to base our intelligent machine upon some particular elaborate, thesaurus-like classification of knowledge, some ad hoc synopticon. Surely that is no road to 'general intelligence.' " 8 And, indeed, little progress has been made toward solving the large data base problem. But, in spite of his own excellent objections, Minsky characteristically concludes: But we had better be cautious about this caution itself, for it exposes us to a far more deadly temptation: to seek a fountain of pure intelligence. I see no reason to believe that intelligence can exist apart from a highly organized body of What Computers Can't Do / 722 knowledge, models, and processes. The habit of our culture has always been to suppose that intelligence resides in some separated crystalline element, call it consciousness, apprehension, insight, gestalt, or what you will but this is merely to confound naming the problem with solving it. The problem-solving abilities of a highly intelligent person lies partly in his superior heuristics for managing his knowledge-structure and partly in the structure itself; these are probably somewhat inseparable. In any case, there is no reason to suppose that you can be intelligent except through the use of an adequate, particular, knowledge or model structure. 9 But this is no argument for optimism. True, people manage to be intelligent, but without the ontological assumption this would be no consolation to workers in AI. It is by no means obvious that in order to be intelligent human beings have somehow solved or needed to solve the large data base problem. The problem may itself be an artifact created by the fact that the computer must operate with discrete elements. Human knowledge does not seem to be analyzable into simple categories as Minsky would like to believe. A mistake, a collision, an embarrassing situation, etc., do not seem on the face of it to be objects or facts about objects. Even a chair is not understandable in terms of any set of facts or "elements of knowledge." To recognize an object as a chair, for example, means to understand its relation to other objects and to human beings. This involves a whole context of human activity of which the shape of our body, the institution of furniture, the inevitability of fatigue, constitute only a small part. And these factors in turn are no more isolable than is the chair. They all may get their meaning in the context of human activity of which they form a part (see Chapter 8). In general, we have an implicit understanding of the human situation which provides the context in which we encounter specific facts and make them explicit. There is no reason, only an ontological commitment, which makes us suppose that all the facts we can make explicit about our situation are already unconsciously explicit in a "model structure," or that we could ever make our situation completely explicit even if we tried. 10 * Why does this assumption seem self-evident to Minsky? Why is he so unaware of the alternative that he takes the view that intelligence in- The Ontological Assumption / 723 volves a "particular, knowledge or model structure," a great systematic array of facts, as an axiom rather than as an hypothesis? Ironically, Minsky supposes that in announcing this axiom he is combating the tradition. "The habit of our culture has always been to suppose that intelligence resides in some separated crystalline element, call it con- sciousness, apprehension, insight, gestalt. . . ." In fact, by supposing that the alternatives are either a well-structured body of facts, or some disem- bodied way of dealing with the facts, Minsky is so traditional that he can't even see the fundamental assumption that he shares with the whole of the philosophical tradition. In assuming that what is given are facts at all, Minsky is simply echoing a view which has been developing since Plato and has now become so ingrained as to seem self-evident. As we have seen, the goal of the philosophical tradition embedded in our culture is to eliminate all risk: moral, intellectual, and practical. Indeed, the demand that knowledge be expressed in terms of rules or definitions which can be applied without the risk of interpretation is already present in Plato, as is the belief in simple elements to which the rules apply. 11 * With Leibniz, the connection between the traditional idea of knowledge and the Minsky-like view that the world must be analyz- able into discrete elements becomes explicit. According to Leibniz, in understanding we analyze concepts into more simple elements. In order to avoid a regress of simpler and simpler elements, then, there must be ultimate simples in terms of which all complex concepts can be under- stood. Moreover, if concepts are to apply to the world, there must be logical simples to which these elements apply. Leibniz envisaged "a kind of alphabet of human thoughts" 12 whose "characters must show, when they are used in demonstrations, some kind of connection, grouping and order which are also found in the objects." 13 The empiricist tradition, too, is dominated by the idea of discrete elements of knowledge. For Hume, all experience is made up of impressions: isolable, determinate, atoms of experience. Intellectualist and empiricist schools converge in Russell's logical atomism, and the idea reaches its fullest expression in Wittgen- stein's Tractatus, where the world is defined in terms of a set of atomic facts which can be expressed in logically independent propositions. This is the purest formulation of the ontological assumption, and the neces- What Computers Can't Do / 124 sary precondition of the possibility of AI, given the fact that digital computers, composed of flip/flops, must ultimately contain a model of the world represented as a structured set of facts or propositions which are either true or false. Thus both philosophy and technology finally posit what Plato sought: a world in which the possibility of clarity, certainty, and control is guaranteed; a world of data structures, decision theory, and automation. No sooner had this certainty finally been made fully explicit, how- ever, than philosophers began to call it into question. Continental phenomenologists recognized it as the outcome of the philosophical tradition and tried to show its limitations. Merleau-Ponty calls the as- sumption that all that exists can be treated as a set of atomic facts, the prejuge du monde, "presumption of commonsense." 14 Heidegger calls it rechnende Denken, I5 "calculating thought," and views it as the goal of philosophy, inevitably culminating in technology. Thus, for Heidegger, technology, with its insistence on the "thoroughgoing calculability of objects," 16 * is the inevitable culmination of metaphysics, the exclusive concern with beings (objects) and the concomitant exclusion of Being (very roughly our sense of the human situation which determines what is to count as an object). In England, Wittgenstein less prophetically and more analytically recognized the impossibility of carrying through the ontological analysis proposed in his Tractatus and became his own sever- est critic. 17 * In Part III, we shall have occasion to follow at length the Merleau- Pontyian, Wittgensteinian, and Heideggerian critique of the traditional ontological assumption, and the alternative view they propose. We have already seen enough, however, to suggest that we do not experience the world as a set of facts in our everyday activities, nor is it self-evident that it is possible to carry through such an analysis. But if the ontological assumption does not square with our experience, why does it have such power? Even if what gave impetus to the philo- sophical tradition was the demand that things be clear and simple so that we can understand and control them, if things are not so simple why persist in this optimism? What lends plausibility to this dream? As we have already seen in another connection, the myth is fostered by the The Ontologies! Assumption / 725 success of modern physics. Here, at least to a first approximation, the ontological assumption works. It was only after Galileo was able to treat motion in terms of isolable objects moving under the influence of com- putable, determinate forces that Hobbes was encouraged to announce that all thinking was the addition of parcels. It has proved profitable to think of the physical universe as a set of independent interacting ele- ments. The ontological assumption that the human world too can be treated in terms of a set of elements gains plausibility when one fails to distinguish between world and universe, or what comes to the same thing, between the human situation and the state of a physical system. In Minsky's work this confusion remains implicit; in the work of his former colleague, John McCarthy, now directing AI research at Stan- ford, it becomes the very cornerstone of the argument. In his paper "Programs with Common Sense," included in the Minsky volume, McCarthy proposes an "advice taker" a program for "solving prob- lems by manipulating sentences in formal languages," the behavior of which "will be improvable merely by making statements to it, telling it about its symbolic environment and what is wanted from it." 18 McCarthy sees clearly that "the first requirement for the advice taker is a formal system in which facts about situation, goals, and actions can be ex- pressed." 19 This leads immediately to the basic problem: how can one describe the situation in a formal system? McCarthy, however, does not see this as a serious problem because he assumes without question that a situation is a physical state: One of the basic entities in our theory is the situation. Intuitively, a situation is the complete state of affairs at some instant in time. The laws of motion of a system determine all future situations from a given situation. Thus, a situation corresponds to the notion of a point in phase space. 20 But the same type of situation can reoccur, involving different objects, different people, and a fortiori different physical states. Moreover, the same physical organization of matter can be seen as many different situations, depending on the goals and intentions of the various human beings involved. Thus, although at any given moment the universe is in only one physical state, there may be as many situations as there are What Computers Can't Do / 126 people. When McCarthy says "there is only one situation corresponding to a given value of time,'* 21 he has clearly confused situation with physical state of the universe. More specifically, he has confused token states and types of states. A situation token can be identical with a physical state token (specified by a point in phase space). But a type of situation cannot be identical to a type of physical state. A concrete example will help to pinpoint this confusion. A situation which McCarthy discusses at length is "being at home.'* " 'At (I, home) (s)* means I am at home in situation s." 22 McCarthy seems to assume that this is the same thing as being in my house, that is, that it is a physical state. But I can be at home and be in the backyard, that is, not physically in my house at all. I can also be physically in my house and not be at home; for example, if I own the house but have not yet moved my furniture in. Being at home is a human situation, not in any simple correspondence with the physical state of a human body in a house. Not to mention the fact that it is a necessary if not sufficient condition for being at home in the sense in question that I own or rent the house, and owning or renting a house is a complicated institutional set of relations not reducible to any set of physical states. Even a physical description of a certain pattern of ink deposited on certain pieces of paper in a specific temporal sequence would not constitute a necessary and suffi- cient condition for a transfer of ownership. Writing one's name is not always signing, and watching is not always witnessing. It is easy to see why McCarthy would like to treat the situation as if it were a physical state. The evolution of a physical state can, indeed, be formalized in differential equations and reproduced on a digital com- puter. Situations, however, pose formidable problems for those who would like to translate them into a formal system. Such a formalization may well be impossible in principle, as can best be seen by returning to the problem of machine translation. We have seen in Part I that automatic language translation has failed because natural language turns out to be much more ambiguous than was supposed. In narrowing down this semantic and syntactic ambiguity the native speaker may appeal to specific information about the world. Bar- Hillel makes this point in an argument which according to him "amounts The Ontological Assumption / 727 to an almost full-fledged demonstration of the unattainability of fully automatic high quality translation, not only in the near future but al- together." 23 The argument is sufficiently important at this point to merit quoting at some length. I shall show that there exist extremely simple sentences in English and the same holds, I am sure, for any other natural language which, within certain linguistic contexts, would be uniquely (up to plain synonymy) and unambiguously trans- lated into any other language by anyone with a sufficient knowledge of the two languages involved, though I know of no program that would enable a machine to come up with this unique rendering unless by a completely arbitrary and ad hoc procedure whose futility would show itself in the next example. A sentence of this kind is the following: The box was in the pen. The linguistic context from which this sentence is taken is, say, the following: Little John was looking for his toy box. Finally he found it. The box was in the pen. John was very happy. Assume, for simplicity's sake, that pen in English has only the following two meanings: (1) a certain writing utensil, (2) an enclosure where small children can play. I now claim that no existing or imaginable program will enable an elec- tronic computer to determine that the word pen in the given sentence within the given context has the second of the above meanings, whereas every reader with a sufficient knowledge of English will do this "automatically." What makes an intelligent human reader grasp this meaning so unhesitat- ingly is, in addition to all the other features that have been discussed by MT workers . . . , his knowledge that the relative sizes of pens, in the sense of writing implements, toy boxes, and pens, in the sense of playpens, are such that when someone writes under ordinary circumstances and in something like the given context, "The box was in the pen," he almost certainly refers to a playpen and most certainly not to a writing pen. 24 * And, as Bar-Hillel goes on to argue, the suggestion, such as Minsky's, that a computer used in translating be supplied with a universal ency- clopedia is "utterly chimerical." "The number of facts we human beings know is, in a certain very pregnant sense, infinite." 25 Bar-HillePs point is well taken; his example, however, based on a particular physical fact, is unfortunate; it tempts AI workers such as Minsky to propose a solution in terms of a model of the facts of physics: What Computers Can't Do / 128 ". . . it would be a good idea to build into the semantic model enough common-sense geometrical physics to make it unlikely that the box is in the fountain-pen. . . ." 26 * There is a second kind of disambiguation, however, which gets us to the very heart of the difficulty. In disambiguating, one may appeal to a sense of the situation as in the following example from Katz and Fodor: An ambiguous sentence such as "He follows Marx" occurring in a setting in which it is clear that the speaker is remarking about intellectual history cannot bear the reading "he dogs the footsteps of Groucho." 27 Katz and Fodor discuss this sort of difficulty in their article "The Structure of a Semantic Theory": Since a complete theory of setting selection must represent as part of the setting of an utterance any and every feature of the world which speakers need in order to determine the preferred reading of that utterance, and since . . . practically any item of information about the world is essential to some disambiguations, two conclusions follow. First, such a theory cannot in principle distinguish between the speaker's knowledge of his language and his knowledge of the world. . . . Second, since there is no serious possibility of systematizing all the knowledge about the world that speakers share . . . [such a theory] is not a serious model for linguistics. 25 Katz and Fodor continue: None of these considerations is intended to rule out the possibility that, by placing relatively strong limitations on the information about the world that a theory can represent in the characterization of a setting, a limited theory of selection by sociophysical setting can be constructed. What these considerations do show is that a complete theory of this kind is not a possibility. 29 Thus Bar-Hillel claims we must appeal to specific facts, such as the size of pens and boxes; Katz and Fodor assume we must appeal to the sociophysical setting. The appeal to context, would, moreover, seem to be more fundamental than the appeal to facts, for the context determines the significance of the facts. Thus in spite of our general knowledge about the relative size of pens and boxes, we might interpret "The box is in the pen," when whispered in a James Bond movie, as meaning just the opposite of what it means at home or on the farm. And, conversely, when The Ontological Assumption / 729 no specifically odd context is specified, we assume a "normal" context and assign to the facts about relative size a "normal" significance. Min- sky's physical model hides but does not obviate the need for this implicit appeal to the situation. The important difference between disambiguation by facts and disam- biguation by appeal to the situation is not noted by Minsky, Bar-Hillel, or Fodor and Katz, presumably because they each assume that the setting is itself identified by features which are facts, and functions like a fact in disambiguation. We shall see, however, that disregarding the difference between fact and situation leads to an equivocation in both Bar-Hillel and Fodor-Katz as to whether mechanical translation is im- practical or impossible. In Bar-HilleFs "demonstration" that since disambiguation depends on the use of facts, and the number of facts is "in a certain very pregnant sense infinite," fully automatic high-quality mechanical translation is unattainable; it is unclear what is being claimed. If "unattainable" means that in terms of present computers, and programs in operation or en- visaged, no such massive storage and retrieval of information can be carried out, then the point is well made, and is sufficient to cast serious doubt on claims that mechanical translation has been achieved or can be achieved in the foreseeable future. But if "unattainable" means theoreti- cally impossible which the appeal to infinity seems to imply then Bar-Hillel is claiming too much. A machine would not have to store an infinite number of facts, for, as Minsky sees, from a large number of facts and rules for concatenating them, such as the laws of physics, it could produce further ones indefinitely. True, no present program would en- able a machine to sort through such an endless amount of data. At present there exist no machine and no program capable of storing even a very large body of data so as to gain access to the relevant information in manageable time. Still, there is work being done on what are called "associative memories" and ingenious tricks used in programming, such as hash coding, which may in the distant future provide the means of storing and accessing vast bodies of information. Then if all that was needed was facts, the necessary information might be stored in such a What Computers Can't Do / 130 way that in any given case only a finite number of relevant facts need be considered. As long as Katz and Fodor, like Bar-Hillel, accept the ontological assumption and speak of the setting in terms of "items of information, 5 * their argument is as equivocal as his. They have no right to pass from the claim that there is "no serious possibility" of systematizing the knowledge necessary for disambiguation, which seems to be a statement about our technological capabilities, to the claim that a complete theory of selection by sociophysical setting is "not a possibility." If a program for handling all knowledge is ever developed, and in their world there is no theoretical reason why it should not be, it will be such a theory. Only if one rejects the ontological assumption that the world can be analyzed as a set of facts items of information can one legitimately move beyond practical impossibility. We have already seen examples which suggest that the situation might be of a radically different order and fulfill a totally different function than any concatenation of facts. In the "Marx" example, the situation (academic) determines how to disam- biguate "Marx" (Karl) and furthermore tells us which facts are relevant to disambiguate "follows," as ideological or chronological. (When was the follower born, what are his political views, etc.?) In the box-pen example the size of the box and pen are clearly relevant since we are speaking of physical objects being "in" other physical objects; but here the situation, be it agricultural, domestic, or conspiratorial, determines the significance of the facts involved. Thus it is our sense of the situation which enables us to select from the potential infinity of facts the immedi- ately relevant ones, and once these relevant facts are found, enables us to estimate their significance. This suggests that unless there are some facts whose relevance and significance are invariant in all situations and no one has come up with such facts we will have to give the computer a way of recognizing situations; otherwise, it will not be able to disambiguate and thus it will be, in principle, unable to understand utterances in a natural language. Among workers in AI, only Joseph Weizenbaum seems to be aware of these problems. In his work on a program which would allow people to converse with a computer in a natural language, Weizenbaum has had The Ontological Assumption / 131 to face the importance of the situation, and realizes that it cannot be treated simply as a set of facts. His remarks on the importance of global context are worth quoting at length: No understanding is possible in the absence of an established global context. To be sure, strangers do meet, converse, and immediately understand one another. But they operate in a shared culture provided partially by the very language they speak and, under any but the most trivial circumstances, engage in a kind of hunting behavior which has as its object the creation of a contextual frame- work. 30 In real conversation global context assigns meaning to what is being said in only the most general way. The conversation proceeds by establishing subcon- texts, sub-subcontexts within these, and so on. 31 Weizenbaum sees difficulties in all this but no problems of principle. I call attention to the contextual matter ... to underline the thesis that, while a computer program that "understands*' natural language in the most general sense is for the present beyond our means, the granting of even a quite broad contextual framework allows us to construct practical language recognition procedures. 32 Thus, Weizenbaum proposes to program a nest of contexts in terms of a "contextual tree": "beginning with the topmost or initial node, a new node representing a subcontext is generated, and from this one a new node still, and so on to many levels.'* 33 He clearly supposes these contexts can themselves ultimately be treated as sets of facts: "the analogue of a conversation tree is what the social psychologist Abelson calls a brief structure," 34 that is, an organized collection of facts concerning a per- son's knowledge, emotional attitudes, goals, and so forth. Evidently, an understanding of the crucial role of the situation does not by itself constitute a sufficient argument for abandoning AI. The traditional ontologist, reincarnated in Weizenbaum and every AI re- searcher, can grant that facts used in conversation are selected and interpreted in terms of the global context and simply conclude that we need only first pick out and program the features which identify this broader situation. But Weizenbaum's observations contain the elements of an objection in principle to the development of humanly intelligent machines. To see this we must first show that Weizenbaum's way of What Computers Can't Do / 132 analyzing the problem separating the meaning of the context from the meaning of the words used in the context is not accidental but is dictated by the nature of a digital machine. In our everyday experience we do not find ourselves making such a separation. We seem to under- stand the situation in terms of the meaning of the words as much as we understand the meaning in terms of the situation. For a computer, however, this reciprocal determination must be broken down into a series of separate operations. Since Weizenbaum sees that we cannot determine the sense of the words until we know the meaning of the context, he correctly concludes, from a programmer's point of view, that we must first specify the context and then use this fixed context to determine the meaning of the elements in it. Moreover, Weizenbaum's analysis suggests that the computerized un- derstanding of a natural language requires that the contexts be organized as a nested hierarchy. To understand why Weizenbaum finds it necessary to use a hierarchy of contexts and work down from the top node, we must return to the general problem of situation recognition. If computers must utilize the situation or context in order to disambiguate, and in general to understand utterances in a natural language, the programmer must be able to program into the machine, which is not involved in a situation, a way of recognizing a context and using it. But the same two problems which arose in disambiguation and necessitated appeal to the situation in the first place arise again on the level of context recognition and force us to envisage working down from the broadest context: (1) If in disam- biguation the number of possibly relevant facts is in some sense infinite so that selection criteria must be applied before interpretation can begin, the number of facts that might be relevant to recognizing a context is infinite too. How is the computer to consider all the features such as how many people are present, the temperature, the pressure, the day of the week, and so forth, any one of which might be a defining feature of some context? (2) Even if the program provides rules for determining relevant facts, these facts would be ambiguous, that is, capable of defining several different contexts, until they were interpreted. Evidently, a broader context will have to be used to determine which of the infinity of features is relevant, and how each is to be understood. The Ontological Assumption / 133 But if, in turn, the program must enable the machine to identify the broader context in terms of its relevant features and this is the only way a computer which operates in terms of discrete elements could proceed the programmer must either claim that some features are intrinsically relevant and have a fixed meaning regardless of context a possibility already excluded in the original appeal to context or the programmer will be faced with an infinite regress of contexts. There seems to be only one way out: rather than work up the tree to ever broader contexts the computer must work down from an ultimate context what Weizen- baum calls our shared culture. Fortunately, there does seem to be something like an ultimate context, but, as we shall see, this proves to be as unprogrammable as the regress it was introduced to avoid. We have seen that in order to identify which facts are relevant for recognizing an academic or a conspiratorial situa- tion, and to interpret these facts, one must appeal to a broader context. Thus it is only in the broader context of social intercourse that we see we must normally take into account what people are wearing and what they are doing, but not how many insects there are in the room or the cloud formations at noon or a minute later. Also only this broader context enables us to determine whether these facts will have their nor- mal significance. Moreover, even the facts necessary to recognize social intercourse can only be singled out because social intercourse is a subcase of human activity, which also includes working alone or studying a primitive tribe. And finally, human activity itself is only a subclass of some even broader situation call it the human life-world which would have to include even those situations where no human beings were directly involved. But what facts would be relevant to recognizing this broadest situation? Or does it make sense to speak of "recognizing" the life-world at all? It seems we simply take for granted this ultimate situation in being people. As Wittgenstein puts it: What has to be accepted, the given is someone could say forms of life." Well then, why not make explicit the significant features of the human form of life from within it? Indeed, this deus ex machina solution has What Computers Can't Do / 134 been the implicit goal of philosophers for two thousand years, and it should be no surprise that nothing short of a formalization of the human form of life could give us artificial intelligence (which is not to say that this is what gives us normal intelligence). But how are we to proceed? Everything we experience in some way, immediate or remote, reflects our human concerns. Without some particular interest, without some partic- ular inquiry to help us select and interpret, we are back confronting the infinity of meaningless facts we were trying to avoid. It seems that given the artificial intelligence worker's conception of reason as calculation on facts, and his admission that which facts are relevant and significant is not just given but is context determined, his attempt to produce intelligent behavior leads to an antinomy. On the one hand, we have the thesis: there must always be a broader context; other- wise, we have no way to distinguish relevant from irrelevant facts. On the other hand, we have the antithesis: there must be an ultimate context, which requires no interpretation; otherwise, there will be an infinite regress of contexts, and we can never begin our formalization. Human beings seem to embody a third possibility which would offer a way out of this dilemma. Instead of a hierarchy of contexts, the present situation is recognized as a continuation or modification of the previous one. Thus we carry over from the immediate past a set of anticipations based on what was relevant and important a moment ago. This carryover gives us certain predispositions as to what is worth noticing. Programming this alternative, however, far from solving the problem of context recognition merely transforms a hierarchical regress into a temporal one. How does the situation which human beings carry along get started? To the programmer this becomes the question: how can we originally select from the infinity of facts those relevant to the human form of life so as to determine a context we can sequentially update? Here the answer seems to be: human beings are simply wired genetically as babies to respond to certain features of the environment such as nipples and smiles which are crucially important for survival. Programming these initial reflexes and letting the computer learn might be a way out of the context recognition problem; but it is important to note two reservations: no present work in artificial intelligence is devoted to this The Ontological Assumption / 135 approach. 36 * In fact, artificial intelligence as it is now defined by Feigen- baum, Simon, Minsky, Weizenbaum, and others seems to be the attempt to produce fully formed adult intelligence, the way Athena sprang full grown from the head of Zeus. Moreover, it is by no means clear that the above proposal avoids the original dilemma. It leaves unexplained how the child develops from fixed responses elicited by fixed features of the environment, to the determination of meaning in terms of context which even AI workers agree characterizes the adult. Once the child can determine meanings in terms of the situation, the past situation can indeed be updated to arrive at the present one, but the original transition from fixed response to flexible response in terms of the meaning of the situation remains as obscure as before. Either the transi- tion must be understood as an ongoing modification of the previous situation, and we have assumed what was to be explained, or the so- called global context must be recognized in terms of fixed context-free features, and we have ignored the problem rather than solved it. Either the child or machine is able to select relevant facts, assign a normal significance to all relevant facts, and also to override this normal signifi- cance in an open-ended way and then no set of fixed features, not even the infant's, can be taken as having a fixed significance in terms of which to begin this process; or fixed features are all that is needed, but then we have to reject as illusory the very flexibility we were trying to explain. There seems to be no way to get into a situation and no way to recognize one from the outside. We nonetheless observe that generality and flexibility are developed gradually through learning, but now the whole problem is hidden in this learning process. The child seems at each moment to be either developing more complex fixed responses, or to have always already interpreted specific facts in terms of the overall context and to be gaining a more structured sense of the situation. If we reject the analysis in terms of fixed responses as inadequate because inapplicable to the adult, we are back facing a temporal version of the original antinomy. Either there must be a first context which a machine would not be able to recognize for want of a previous context in terms of which to single out its relevant features, or there will be a temporal regress of contexts extending infinitely into What Computers Can't Do / 136 the past and the machine will not be able to begin the recognition process. As Kant noted, the resolution of an antinomy requires giving up the assumption that the two alternatives considered are the only possible ones. They are, indeed, the only alternatives open to someone trying to construct artificial reason. 37 * There must be another alternative, how- ever, since language is used and understood. There must be some way of avoiding the self-contradictory regress of contexts, or the incompre- hensible notion of recognizing an ultimate context, as the only way of giving significance to independent, neutral facts. The only way out seems to be to deny the separation of fact and situation, which we saw Weizen- baum was led to assume because of the serial procedure forced on him by the digital computer. If, as all agree, we are unable to eliminate the situation in favor of facts whose relevance and significance are fixed regardless of context, then the only alternative way of denying the sepa- ration of fact and situation is to give up the independence of the facts and understand them as a product of the situation. This would amount to arguing that only in terms of situationally determined relevance are there any facts at all. It also amounts to avoiding the problem of how to recognize the situation from outside by arguing that for an intelligence to have any facts to interpret, it must already be in a situation. Part III will show how this latter alternative is possible and how it is related to the rest of human life. Only then will it become clear why the fixed-feature alternative is empirically untenable, and also why the hu- man form of life cannot be programmed. Conclusion In surveying the four assumptions underlying the optimistic interpreta- tion of results in AI we have observed a recurrent pattern: In each case the assumption was taken to be self-evident an axiom seldom ar- ticulated and never called into question. In fact, the assumption turned out to be only one alternative hypothesis, and a questionable one at that. The biological assumption that the brain must function like a digital computer no longer fits the evidence. The others lead to conceptual difficulties. The psychological assumption that the mind must obey a heuristic program cannot be defended on empirical grounds, and a priori argu- ments in its defense fail to introduce a coherent level of discourse be- tween the physical and the phenomenological. This does not show that the task set for Cognitive Simulation is hopeless. However, this lack of defense of the psychological axiom does eliminate the only argument which suggested any particular reason for hope. If it could have been argued that information processing must proceed by heuristic rules, Cognitive Simulation would have had the promising task of finding these rules. Without the defense provided by this axiom, however, all difficul- ties besetting Cognitive Simulation research during the past ten years take on new significance; there is no reason to deny the growing body of evidence that human and mechanical information processing proceed in entirely different ways. 7/37 What Computers Can't Do / 138 Researchers in AI (taking over from CS as Minsky has taken over from Simon) have written programs which allow the digital machine to approximate, by means of logical operations, the result which human beings seem to achieve by avoiding rather than resolving the difficulties inherent in formalization. But formalization of restricted contexts is an ad hoc "solution" which leaves untouched the problem of how to formal- ize the totality of human knowledge presupposed in intelligent behavior. This fundamental difficulty is hidden by the epistemological and ontolog- ical assumptions that all human behavior must be analyzable in terms of rules relating atomic facts. But the conceptual difficulties introduced by these assumptions are even more serious than those introduced by the psychological one. The inevitable appeal to these assumptions as a final basis for a theory of practice leads to a regress of more and more specific rules for applying rules or of more and more general contexts for recognizing contexts. In the face of these contradictions, it seems reasonable to claim that, on the information processing level, as opposed to the level of the laws of physics, we cannot analyze human behavior in terms of rule-governed manipulation of a set of elements. And since we have seen no argument brought forward by the AI theorists for the assumption that human behavior must be reproducible by a digital computer operating with strict rules on determinate bits, we would seem to have good philosoph- ical grounds for rejecting this assumption. If we do abandon all four assumptions, then the empirical data avail- able to date would take on different significance. It no longer seems obvious that one can introduce search heuristics which enable the speed and accuracy of computers to bludgeon through in those areas where human beings use more elegant techniques. Lacking any a priori basis for confidence, we can only turn to the empirical results obtained thus far. That brute force can succeed to some extent is demonstrated by the early work in the field. The present difficulties in game playing, language translation, problem solving, and pattern recognition, however, indicate a limit to our ability to substitute one kind of "information processing** for another. Only experimentation can determine the extent to which newer and faster machines, better programming languages, and cleverer Conclusion / 139 heuristics can continue to push back the frontier. Nonetheless, the dra- matic slowdown in the fields we have considered and the general failure to fulfill earlier predictions suggest the boundary may be near. Without the four assumptions to fall back on, current stagnation should be grounds for pessimism. This, of course, has profound implications for our philosophical tradi- tion. If the persistent difficulties which have plagued all areas of artificial intelligence are reinterpreted as failures, these failures must be interpre- ted as empirical evidence against the psychological, epistemological, and ontological assumptions. In Heideggerian terms this is to say that if Western Metaphysics reaches its culmination in Cybernetics, the recent difficulties in artificial intelligence, rather than reflecting technological limitations, may reveal the limitations of technology. PART III ALTERNATIVES TO THE TRADITIONAL ASSUMPTIONS Introduction The psychological, epistemological, and ontological assumptions have this in common: they assume that man must be a device which calculates according to rules on data which take the form of atomic facts. Such a view is the tidal wave produced by the confluence of two powerful streams: first, the Platonic reduction of all reasoning to explicit rules and the world to atomic facts to which alone such rules could be applied without the risks of interpretation; second, the invention of the digital computer, a general-purpose information-processing device, which cal- culates according to explicit rules and takes in data in terms of atomic elements logically independent of one another. In some other culture, the digital computer would most likely have seemed an unpromising model for the creation of artificial reason, but in our tradition the computer seems to be the very paradigm of logical intelligence, merely awaiting the proper program to accede to man's essential attribute of rationality. The impetus gained by the mutual reinforcement of two thousand years of tradition and its product, the most powerful device ever invented by man, is simply too great to be arrested, deflected, or even fully understood. The most that can be hoped is that we become aware that the direction this impetus has taken, while unavoidable, is not the only possible direction; that the assumptions underlying the conviction that artificial reason is possible are assumptions, not axioms in short, that What Computers Can't Do / 144 there may be an alternative way of understanding human reason which explains both why the computer paradigm is irresistible and why it must fail. Such an alternative view has many hurdles to overcome. The greatest of these is that it cannot be presented as an alternative scientific explana- tion. We have seen that what counts as "a complete description" or an explanation is determined by the very tradition to which we are seeking an alternative. We will not have understood an ability, such as the human mastery of a natural language, until we have found a theory, a formal system of rules, for describing this competence. We will not have under- stood behavior, such as the use of language, until we can specify that behavior in terms of unique and precisely definable reactions to precisely defined objects in universally defined situations. Thus, Western thought has already committed itself to what would count as an explanation of human behavior. It must be a theory of practice, which treats man as a device, an object responding to the influence of other objects, according to universal laws or rules. But it is just this sort of theory, which, after two thousand years of refinement, has become sufficiently problematic to be rejected by philoso- phers both in the Anglo-American tradition and on the Continent. It is just this theory which has run up against a stone wall in research in artificial intelligence. It is not some specific explanation, then, that has failed, but the whole conceptual framework which assumes that an expla- nation of human behavior can and must take the Platonic form, success- ful in physical explanation; that situations can be treated like physical states; that the human world can be treated like the physical universe. If this whole approach has failed, then in proposing an alternative ac- count we shall have to propose a different sort of explanation, a different sort of answer to the question "How does man produce intelligent behav- ior?" or even a different sort of question, for the notion of "producing" behavior instead of simply exhibiting it is already colored by the tradi- tion. For a product must be produced in some way; and if it isn't produced in some definite way, the only alternative seems to be that it is produced magically. There is a kind of answer to this question which is not committed Introduction / 145 beforehand to finding the precise rulelike relations between precisely defined objects. It takes the form of a phenomenological description of the behavior involved. It, too, can give us understanding if it is able to find the general characteristics of such behavior: what, if any one thing, is involved in seeing a table or a house, or, more generally, in perception, problem solving, using a language, and so forth. Such an account can even be called an explanation if it goes further and tries to find the fundamental features of human activity which serve as the necessary and sufficient conditions for all forms of human behavior. Such an explanation owes a debt to Aristotle's method, although not to his arguments or descriptions. Whereas Plato sought rulelike criteria, Aristotle tried to describe the general structure of perception and judg- ment. But, as his notion that action is based on a practical syllogism shows, Aristotle still thought of man as a calculable and calculating sort of object a reckoning animal so that his actual descriptions are one step in the tradition which finally separated the rationality from the animality and tried to simulate the reckoning all by itself. It is only recently, now that the full implications of the attempt to treat man merely as an object or device have become apparent, that philoso- phers have begun to work out a new view. The pioneers were Heidegger and Wittgenstein. Since then many others, notably Maurice Merleau- Ponty and Michael Polanyi have, each on his own, applied, consolidated, and refined similar insights; and young thinkers such as Charles Taylor and Samuel Todes are continuing their research. In trying to lay out the alternative view that emerges when we confront the three basic assump- tions of the tradition with a phenomenological description of the struc- ture of human behavior, I shall be drawing on the work of all these men. I am fully aware that this "account" is vaguer and less experimental than that of either the behaviorists or intellectualists which it is meant to supplant. But one must not become so fascinated with the formalizable aspects of a subject that one forgets the significant questions which originally gave rise to the research, nor should one be so eager for experimental results that one continues to use old techniques just because they work, when they have ceased to lead to new insights. Chomsky is one of the few in the behavioral sciences who see this danger. What Computers Can't Do / 146 Without wishing to exalt the cult of gentlemanly amateurism, one must neverthe- less recognize that the classical issues have a liveliness and significance that may be lacking in an area of investigation that is determined by the applicability of certain tools and methods, rather than by problems that are of intrinsic interest in themselves. The moral is not to abandon useful tools; rather, it is, first, that one should maintain enough perspective to be able to detect the arrival of that inevitable day when the research that can be conducted with these tools is no longer important; and, second, that one should value ideas and insights that are to the point, though perhaps premature and vague and not productive of research at a particular stage of technique and understanding. 1 Taking this suggestion to heart, we shall explore three areas neces- sarily neglected in CS and AI but which seem to underlie all intelligent behavior: the role of the body in organizing and unifying our experience of objects, the role of the situation in providing a background against which behavior can be orderly without being rulelike, and finally the role of human purposes and needs in organizing the situation so that objects are recognized as relevant and accessible. The Role of the Body in Intelligent Behavior Adherents of the psychological and epistemological assumptions that human behavior must be formalizable in terms of a heuristic program for a digital computer are forced to develop a theory of intelligent behav- ior which makes no appeal to the fact that a man has a body, since at this stage at least the computer clearly hasn't one. In thinking that the body can be dispensed with, these thinkers again follow the tradition, which from Plato to Descartes has thought of the body as getting in the way of intelligence and reason, rather than being in any way indispens- able for it. If the body turns out to be indispensable for intelligent behavior, then we shall have to ask whether the body can be simulated on a heuristically programmed digital computer. If not, then the project of artificial intelligence is doomed from the start. These are the questions to which we must now turn. Descartes, the first to conceive the possibility of robots, was also the first to suggest the essential inadequacy of a finite state machine. He remarks in the Discourses: Although such machines could do many things as well as, or perhaps even better" than men, they would infallibly fail in certain others For while reason is a universal instrument which can be used in all sorts of situations, the organs of a machine have to be arranged in a particular way for each particular action. From this it follows that it is morally [i.e., practically] impossible that there What Computers Can't Do / 148 should be enough different devices in a machine to make it behave in all the occurrences of life as our reason makes us behave. 1 Thus, although not aware of the difference between a situation and a physical state, Descartes already saw that the mind can cope with an indefinite number of situations, whereas a machine has only a limited set of states and so will eventually reveal itself by its failure to respond appropriately. This intrinsic limitation of mechanism, Descartes claims, shows the necessity of presupposing an immaterial soul This is an interesting argument, and some version of it may indeed be valid, but it gets its plausibility from the assumption that a robot can be in only a relatively small number of states. When in a modern computer the number of possible states is of the order of 10 Iol , it is not clear just how much Descartes' objection proves. Such a machine could at least in principle respond to what would appear to be an indefinite number of situations. It would thus, on Descartes' view, be indistinguishable from a human being, destroying his argument that intelligent behavior is possible only if the mechanism behaving is somehow attached to a non- material soul. But one can raise a new objection, in some ways the exact opposite of Descartes'. A brain in a bottle or a digital computer might still not be able to respond to new sorts of situations because our ability to be in a situation might depend, not just on the flexibility of our nervous system, but rather on our ability to engage in practical activity. After some attempts to program such a machine, it might become apparent that what distinguishes persons from machines, no matter how cleverly constructed, is not a detached, universal, immaterial soul but an in- volved, self-moving, material body. Indeed, it is just the bodily side of intelligent behavior which has caused the most trouble for artificial intelligence. Simon, who has been only slightly daunted by the failures of the last ten years, now feels that "machines will be capable, within twenty years, of doing any work that a man can do," 2 but he admits: "Automation of a flexible central nervous system will be feasible long before automation of a comparatively flexible sensory, manipulative, or locomotive system." 3 But what if the work of the central nervous system depends on the locomotive system, or to put The Role of the Body in Intelligent Behavior / 149 it phenomenologically, what if the "higher," determinate, logical, and detached forms of intelligence are necessarily derived from and guided by global and involved "lower" forms? Then Simon's optimism, based on the three assumptions underlying artificial intelligence and traditional philosophy, would be unjustified. The intractability of the "lower" functions has already produced a certain irony. Computer technology has been most successful in simulat- ing the so-called higher rational functions those which were once sup- posed to be uniquely human. Computers can deal brilliantly with ideal languages and abstract logical relations. It turns out that it is the sort of intelligence which we share with animals, such as pattern recognition (along with the use of language, which may indeed be uniquely human) that has resisted machine simulation. Let us reconsider two related areas in which work in artificial intelli- gence has not fulfilled early expectations: game playing and pattern recognition. Thus far I have tried to account for the failure by arguing that the task in question cannot be formalized, and by isolating the nonformal form of "information processing" necessarily involved. Now I shall try to show that the nonformalizable form of "information pro- cessing" in question is possible only for embodied beings. To make this clear we shall first have to consider human pattern recognition in more detail. With the aid of concepts borrowed from phenomenology, I shall try to show how pattern recognition requires a .. certain sort of indeterminate, global anticipation. This set or anticipation is characteristic of our body as a "machine" of nerves and muscles whose function can be studied by the anatomist, and also of our body as ex- perienced by us, as our power to move and manipulate objects in the world. I shall argue that a body in both these senses cannot be repro- duced by a heuristically programmed digital computer even one on wheels which can operate manipulators, and that, therefore, by virtue of being embodied, we can perform tasks beyond the capacities of any heuristically programmed robot. We have seen that the restricted applicability of pattern recognition programs suggests that human pattern recognition proceeds in some What Computers Can't Do / 150 other way than searching through lists of traits. Indeed, phenomenolo- gists and Gestalt psychologists have pointed out that our recognition of ordinary spatial or temporal objects does not seem to operate by check- ing off a list of isolable, neutral, specific characteristics at all. For exam- ple, in recognizing a melody, the notes get their values by being perceived as part of the melody, rather than the melody's being recognized in terms of independently identified notes. Likewise, in the perception of objects there are no neutral traits. The same hazy layer which I would see as dust if I thought I was confronting a wax apple might appear as moisture if I thought I was seeing one that was fresh. The significance of the details and indeed their very look is determined by my perception of the whole. The recognition of spoken language offers the most striking demon- stration of this global character of our experience. From time to time brash predictions such as Rosenblatt's have been made about mechanical secretaries into which (or at whom) one could speak, and whose pro- grams would analyze the sounds into words and type out the results. In fact, no one knows how to begin to make such a versatile device, and further progress is unlikely, for current work has shown that the same physical constellation of sound waves is heard as quite different pho- nemes, depending on the expected meaning. Oettinger has given considerable attention to the problem. His analysis of speech recognition work is worth reproducing in detail, both because this pattern recognition problem is important in itself and because this work exhibits the early success and subsequent failure to generalize which we have come to recognize as typical of artificial intelligence research. There was considerable initial success in building apparatus that would eke out a sequence of discrete phonemes out of the continuous speech waveform. While phonemic analysis has been dominant in that area, numerous other approaches to this decoding problem have also been followed. All have shared this initial degree of success and yet all, so far, have proved to be incapable- of significant expansion beyond the recognition of the speech of a very few distinct individuals and the recognition of a very few distinct sound patterns whether they be pho- nemes or words or whatever. All is well as long as you are willing to have a fairly restricted universe of speakers, or sounds, or of both. Within these limitations you can play some very good tricks. There are now lots of machines, some experimental, some not so experimental, that will recog- The Role of the Body in Intelligent Behavior / 757 nize somewhere between 20 and 100 distinct sound patterns, some of them quite elaborate. Usually the trick is something like identifying a number of features, treating these as if they were coordinates in some hyperspace, then passing planes that cordon off, if you will, different blocks of this space. If your speech event falls somewhere within one of these blocks you say that it must have been that sound and you recognize it. This game was fairly successful in the range of twenty to a hundred or so distinct things, but after that, these blocks become so small and clustered so close together that you no longer can achieve any reliable sort of separation. Every- thing goes to pot. 4 This leads Oettinger to a very phenomenoJogical observation: Perhaps ... in perception as well as in conscious scholarly analysis, the phoneme comes after the fact, namely ... it is constructed, if at all, as a consequence of perception not as a step in the process of perception itself. 5 This would mean that the total meaning of a sentence (or a melody or a perceptual object) determines the value to be assigned to the individual elements. Oettinger goes on reluctantly to draw this conclusion: This drives me to the unpopular and possibly unfruitful notion that maybe there is some kind of Gestalt perception going on, that here you are listening to me, and somehow the meaning of what Fm saying comes through to you all of a piece. And it is only a posteriori, and if you really give a damn, that you stop and say, "Now, here was a sentence and the words in it were of such and such type, and maybe here was a rfoun and here was a vowel and that vowel was this phoneme and the sentence is declarative, etc." 6 Phenomenologists, not committed to breaking down the pattern so that it can be recognized by a digital computer, while less appalled, are no less fascinated by the gestalt character of perception. Indeed, it has been systematically studied in their account of perceptual horizons. Two forms of awareness are involved. First there is the basic figure-ground phenomenon, necessary for there to be any perception at all: whatever is prominent in our experience and engages our attention appears on a background which remains more or less indeterminate. This back- ground, which need never have been made determinate, affects the ap- pearance of what is determinate by letting it appear as a unified, bounded figure. In Rubin's famous "Peter-Paul Goblet" (Figure 3), "the contour What Computers Can't Do / 152 which divides figure from ground 'belongs' to the figure only and changes its shape radically if a figure-ground reversal occurs." 7 Thus the figure has specific determinate characteristics, while the background can be characterized only as that-which-is-not-the figure. Figure 3 This indeterminacy plays a crucial role in human perception. Merleau- Ponty points out that most of what we experience must remain in the background so that something can be perceived in the foreground. When Gestalt theory informs us that a figure on a background is the simplest sense-datum available to us, we reply that this is not a contingent characteriza- tion of factual perception, which leaves us free, in an ideal analysis, to bring in the notion of impression. It is the very definition of the phenomenon of percep- tion. . . . The perceptual 'something* is always in the middle of something else; it always forms part of a 'field.' 8 It is this ground, or outer horizon as Edmund Husserl, the founder of phenomenology, called it, which in our chess example remains indeter- minate and yet provides the context of the specific counting out, so that one always has a sense of the relevance of the specific move under The Role of the Body in Intelligent Behavior / 753 consideration to the rest of the game. Similarly, our sense of the overall context may organize and direct our perception of the details when we understand a sentence. For a computer, which must take up every bit of information explicitly or not at all, there could be no outer horizon. Any information to be taken into account would have to be as determinate as the figure. This leads to the unwieldy calculations which we have seen in chess programs and which Oettinger deplores in language pro- grams. This outer horizon, then, describes how background "information" about a conversation or a particular game is ignored without being excluded. It does not, however, describe the way the background pro- vides information which contributes to the player zeroing in on one area of the chess board rather than another, or how our anticipation of a sentence's meaning determines our understanding of its elements as they fall into place. To understand this, we must consider a second kind of perceptual indeterminacy investigated by Husserl and Gestalt psycholo- gists: what Husserl calls the inner horizon. The something-more-than- the-figure is, in this case, not as indeterminate as the outer horizon. When we perceive an object we are aware that it has more aspects than we are at the moment considering. Moreover, once we have experienced these further aspects, they will be experienced as copresent, as covered up by what is directly presented. Thus, in ordinary situations, we say we perceive the whole object, even its hidden aspects, because the concealed aspects directly affect our perception. We perceive a house, for example, as more than a facade as having some sort of back some inner hori- zon. We respond to this whole object first and then, as we get to know the object better, fill in the details as to inside and back. A machine with no equivalent of an inner horizon would have to process this information in the reverse order: from details to the whole. Given any aspect of an object, the machine would either pick it up on its receptors or it would not. All additional information about other aspects of the object would have to be explicitly stored in memory in Minsky's sort of model or counted out again when it was needed. This lack of horizons is the essential difference between an image in a movie or on a TV screen and the same scene as experienced by a human being. What Computers Can't Do / 154 When, in a film, the camera is trained on an object and moves nearer to it to give a close-up view, we can remember that we are being shown the ash tray or an actor's hand, we do not actually identify it. This is because the scene has no horizons. 9 In chess and in recognizing sentences, we find the same phenomenon playing a crucial role. Our sense of the whole situation, outer horizon, and our past experience with the specific object or pattern in question, inner horizon, give us a sense of the whole and guide us in filling in the details. 10 * This process can best be noticed when it is breaking down. If you reach for a glass of water and get milk by mistake, on taking a sip your first reaction is total disorientation. You don't taste water, but you don't taste milk either. You have a mouthful that approaches what Husserl would call pure sensuous matter or hyletic data, and naturally you want to spit it out. Or, if you find the right global meaning fast enough, you may recover in time to recognize the milk for what it is. Its other characteris- tics, whether it is fresh or sour, buttermilk or skimmed milk, will then fall into place. One might well wonder how one knows enough to try "milk" rathe'r than, say, "gasoline.*' Doesn't one need some neutral features to begin this process of recognition? The perceiver's apparent clairvoyance seems so paradoxical that one is tempted to embrace the computer model in spite of its difficulties. But the process seems less mysterious when we bear in mind that each new meaning is given in an outer horizon which is already organized, in this case a meal, on the basis of which we already have certain expectations. It is also important that we sometimes do give the wrong meaning; in these cases the data coming in make no sense at all, and we have to try a new total hypothesis. A computer, which must operate on completely determinate data according to strictly defined rules, could at best be programmed to try out a series of hypotheses to see which best fit the fixed data. But this is far from the flexible interaction of underdetermined data and underde- termmed expectations which seems to be characteristic of human pattern recognition. As one might expect, the computer people, again with the support of The Role of the Body in Intelligent Behavior / 755 the philosophical tradition, and the success of physics, have rarely faced this problem. Philosophers have thought of man as a contemplative mind passively receiving data about the world and then ordering the elements. Physics has made this conception plausible on the level of the brain as a physical object. The brain does passively receive energy from the physical world and process it in terms of its present state which is a function of past energy received. If one accepts the passive view of mind and fails to distinguish the physical-processing level from the "informa- tion-processing" level, it seems self-evident that the mind, like the com- puter, simply receives bits of determinate data. In his introduction to the Scientific American issue on computers, McCarthy naively confuses brain and mind, energy and information, so that the passivity of the computer appears to be a self-evident model for human "information processing." The human brain also accepts inputs of information, combines it with informa- tion stored somehow within, and returns outputs of information to its environ- ment. 11 . Neisser is much more subtle. He too underestimates the problems posed by the role of anticipation, but his work in psychology has at least led him to see the need for "wholistic operations which form the units to which attention may then be directed," 12 and he tries to fit this fact into his overall commitment to a digital computer model. The result is a confusion between what "global or wholistic" means in a gestalt analy- sis and what it would have to mean in a computer program, which is sufficiently revealing to be worth following in detail. A general characterization of the gestalt, or global, phenomenon is: the interpretation of a part depends on the whole in which it is embed- ded. But this is too general. Such a definition allows Minsky, for example, to miss the whole problem. In his Scientific American article he speaks of Evans' analogy-solving program as being able to "recognize a 'global' aspect of the situation." 13 This turns out to mean that, on the basis of calculations made on certain local features of a figure, the program segments two superimposed figures in one way rather than another. There is nothing here to surprise or interest those concerned with the What Computers Can't Do / 156 way the gestalt, or global, configuration functions in our experience. To see the difference between the wholistic processes which interest Neisser and what Minsky calls global recognition, one needs a sharper characterization of the gestalt phenomenon. Neisser gives such a charac- terization in terms of a temporal gestalt, a rhythm (a favorite example of the Gestaltists): The parts (individual beats) get their meaning (relative position) from the whole, even though that whole does not exist at any moment of time. It exists, as one might say, in the subject's mind, as an intent . . . Gestalt. . . . u The crucial feature of this gestalt interpretation, that each part gets its meaning only in terms of the whole, is missing in Minsky's example, as it must be, since, as we have seen, for a digital computer, each complex whole must be constructed by the logical combination of independently defined elements. In Minsky's example, the elements already have a precise significance (or rather two possible precise significances), and it is simply a question of deciding which interpretation is appropriate in terms of a decision based on other determinate local features of the figure. Neisser's description of the "mind's intent" which gives the individual beats their significance, on the other hand, brings us to the center of the problem. The question is how the partially determinate anticipation, involved in game playing, pattern recognition, and intelligent behavior in general, can be simulated on a heuristically programmed digital com- puter so that a computer does not have to go through the enormous calculation required by an explicit internal model. More specifically for Neisser, the problem is how to reconcile his gestaltist analysis with a computer model of human performance. Neisser thinks he has a way. In discussing linguistic performance as an example of the gestalt eifect, Neisser thinks of the rules of grammar as the wholes into which the words fit as parts. The rules are structural That is, they do not dictate what particular words are to be used, but rather how they are to be related to each other and to the sentence as a whole. 15 The Role of the Body in Intelligent Behavior / 757 But this will not work. In the case of the rhythm, the whole determined the meaning of each element there is no such thing as a syncopated beat, for example, existing all by itself but for Neisser, in the case of language, the words already have a determinate set of possible meanings; the grammar simply provides a rule for selecting a meaning and combin- ing it with others. The elements in this case are completely determinate and can be defined independently of the rules. It is, therefore, misleading when Neisser concludes: r A sentence is more than the sum of its parts. This is not an unfamiliar slogan. Long ago, the Gestalt psychologists used it to describe the wholistic aspects of visual perception. 16 This confusion is already latent in Neisser's description of the anticipa- tion involved in hearing a rhythm in the example quoted above. The description concludes: "[The anticipation] exists ... in the subject's mind as an intent, a gestalt, a plan, a description of a response that can be executed without further consideration." 11 This slide from gestalt antici- pation to preset plan is an obfuscation necessitated by the computer model: A gestalt determines the meaning of the elements it organizes; a plan or a rule simply organizes independently defined elements. More- over, just as the elements (the beats) cannot be defined independently of the gestalt, the gestalt (the rhythm) is nothing but the organization of the elements. A plan, on the other hand, can be stated as a rule or program, independently of the elements. Clearly his computer model of a formal program defined and stored separately from the independently defined bits of data which it organizes leads Neisser to betray his own gestaltist illustration. This difference is neglected in all CS models, yet it is the essence of the gestaltist insight, and accounts for the flexibility of human pattern recognition compared to that of machines. Thus far computer programs have been unable to approach this inter- dependence of parts and whole. Neisser himself never sees this problem, but he unwittingly casts some new light on the important differences between mechanist and gestaltist models of psychological processes when he contrasts the digital model of neural processes postulated by the What Computers Can't Do / 158 transformational linguists with the analogue model of the brain espoused by the early Gestalt psychologists. [The Gestaltists] were "nativists," believing that the perceptual processes were determined by necessary and innate principles rather than by learning. The proper figural organization was due to processes in the brain, which followed unvarying (and wholistic) laws of physics and chemistry The perceived world always took the "best," the "structurally simplest" form, because of the equilib- rium principle that transcends any possible effects of learning or practice. 15 Such an analogue model of brain function, in which information is integrated by equilibrium forces rather than on/off switches, was neces- sary if the Gestalt psychologists were to account for the role of global anticipations in structuring experience. They had been led to break with the rationalist tradition running from Descartes to Kant, which con- ceived of the mind as bringing independently defined innate principles (Descartes) or rules (Kant) to bear on otherwise unstructured experi- ence. This rationalist conception (with the addition of minimal bits of determinate experience) lends itself perfectly to a computer model, but the Gestaltists saw that their principles of organization like the equilib- rium patterns formed by charged particles on curved surfaces could not be separated from the elements they organized. Thus, even if the digital model of the brain had existed at the time, the Gestaltists would have rejected it. 19 * Neisser does not see this. He supposes that the digital model of built-in rules, which the linguists have been led to propose, is an improvement on the analogue model proposed by the Gestaltists. Neisser's praise of the linguists' "improvement," ignoring as it does the difficulties in artifi- cial intelligence, the latest developments in neurophysiology, and the reason the Gestaltists proposed an analogue model in the first place can only be a non sequitur: The Gestalt psychologists were never able to provide any satisfactory description or analysis of the structures involved in perception. The few attempts to specify "fields of force" in vision, or "ionic equilibria" in the brain, were ad hoc and ended in failure. In linguistics, by contrast, the study of "syntactic structures" has a long history. 20 The Role of the Body in Intelligent Behavior / 759 How the long history of syntactic structures is supposed to show that the linguists have a better model of neural processes than the Gestaltists is totally unclear. It seems to mean that at least the rules the linguists are looking for would be, if they were found, the sort of rules one could process with a digital computer which we already understand, whereas the gestaltist equilibrium principles could only be simulated on a brain- like analogue computer, which no one at present knows how to design. This is no doubt true, but it reminds one of the story of the drunk who lost a key in the dark but looked for it under a street lamp because the light was better. It would indeed be nice to have a programmable model in linguistics, and in psychology in general, but the fact remains that modern linguists have no more detailed account of what goes on in the brain than did the Gestaltists, and, moreover, as a theory of competence, not performance, modern linguistics is not even trying to provide an- swers to the problem of how we produce intelligent behavior. Worse, in this case, the street lamp is not even lit. We have seen that when digital computers have been used to try to simulate linguistic performance, they have had remarkably little success. The upshot of Neisser's comparison of gestalt and linguistic models of the brain, in opposition to his intent, is to call attention to a difference in brain model which exactly parallels the difference in the conception of the wholistic processes, which he also overlooks. The sort of gestalt process illustrated in Neisser's example of the rhythm which gives mean- ing to and is made up of its beats suggests that however the brain integrates stimuli, it does not do it like a digital computer applying independently defined heuristic rules to independently defined bits of data. Among computer experts only Donald MacKay has seen this point. He concludes: It may well be that only a special-purpose 'analogue* mechanism could meet ail detailed needs We on the circuit side had better be very cautious before we insist that the kind of information processing that a brain does can be replicated in a realizable circuit. Some kind of 'wet' engineering may turn out to be inevi- table. 21 What Computers Can't Do / 160 If, in the light of the phenomenological and neurophysiological evi- dence, we accept the view that the nervous system is some sort of analogue computer operating with equilibrium fields, we must still be on guard against transferring to psychology this model of the nervous sys- tem, conceived as a brain in a bottle receiving energy from the world and sending out responses. The human perceiver must be understood in different terms than his nervous system. To have an alternative account of intelligent behavior we must describe the general and fundamental features of human activity. In the absence of a workable digital computer model, and leaving to the neurophysiologist the question of how the brain integrates incoming physical stimuli, we must again ask, How do human beings use an underdetermined, wholistic expectation to organize their experience? Husserl has no further account beyond the assertion that we do: that "transcendental consciousness" has the ''wunderbar" capacity for giving meanings and thus making possible the perception, recognition, and exploration of enduring objects. Like the Gestaltists, he thinks of these meanings as partially indeterminate wholes, not as explicit programs or rules. But even Husserl is not free from the traditional intellectualist view, and thus he too is vulnerable to the criticism directed at Neisser. Husserl, like Descartes and Kant, thinks of form as separable from content, of the global anticipation as separable from its sensuous feeling. Thus, his noema, or perceptual anticipation, is like a rule or program in one crucial way: it exists in the mind or transcendental consciousness independently of its application to the experience it structures. Merleau-Ponty tries to correct HusserPs account on this point and at the same time develop a general description which supports the Gestalt- ists. He argues that it is the body which confers the meanings discovered by Husserl. After all, it is our body which captures a rhythm. We have a body-set to respond to the sound pattern. This body-set is not a rule in the mind which can be formulated or entertained apart from the actual activity of anticipating the beats. Generally, in acquiring a skill in learning to drive, dance, or pro- nounce a foreign language, for example at first we must slowly, awk- wardly, and consciously follow the rules. But then there comes a moment when we finally transfer control to the body. At this point we do not seem The Role of the Body in Intelligent Behavior / 767 to be simply dropping these same rigid rules into unconsciousness; rather we seem to have picked up the muscular gestalt which gives our behavior a new flexibility and smoothness. The same holds for acquiring the skill of perception. To take one of Merleau-Ponty's examples: to learn to feel silk, one must learn to move or be prepared to move one's hand in a certain way and to have certain expectations. Before we acquire the appropriate skill, we experience only confused sensations. It is easiest to become aware of the body's role in taste, hearing, and touch, but seeing, too, is a skill that has to be learned. Focusing, getting the right perspective, picking out certain details, all involve coordinated actions and anticipations. As Piaget remarks, "Perceptual constancy seems to be the product of genuine actions, which consist of actual or potential movements of the glance or of the organs concerned. . . ." 22 These bodily skills enable us not only to recognize objects in each single sense modality, but by virtue of the felt equivalence of our explora- tory skills we can see and touch the same object. A computer to do the same thing would have to be programmed to make a specific list of the characteristics of a visually analyzed object and compare that list to an explicit list of traits recorded by moving tactical receptors over that same object. This means that there would have to be an internal model of each object in each sense modality, and that the recognition of an object seen and felt must pass through the analysis of that object in terms of common features. My body enables me to by-pass this formal analysis. A skill, unlike a fixed response or set of responses can be brought to bear in an indefinite number of ways. When the percipient acquires a skill, he does not weld together individual movements and individual stimuli but acquires the power to respond with a certain type of solution to situations of a certain general form. The situations may differ widely from place to place, and the response movements may be entrusted sometimes to one operative organ, some- times to another, both situations and responses in the various cases having in common not so much a partial identity of elements as a shared significance. 23 Thus I can recognize the resistance of a rough surface with my hands, with my feet, or even with my gaze. My body is thus what Merleau-Ponty calls a "synergic system,'* 24 "a ready-made system of equivalents and transpositions from one sense to another." 25 What Computers Can't Do / 162 Any object presented to one sense calls upon itself the concordant operation of all the others. I see a surface colour because I have a visual field, and because the arrangement of the field leads my gaze to that surface I perceive a thing because I have a field of existence and because each phenomenon, on its appear- ance, attracts towards that field the whole of my body as a system of perceptual powers. 26 A human perceiver, like a machine, needs feedback to find out if he has successfully recognized an object. But here too there is an important difference in the feedback involved. A machine can, at best, make a specific set of hypotheses and then find out if they have been confirmed or refuted by the data. The body can constantly modify its expectations in terms of a more flexible criterion: as embodied, we need not check for specific characteristics or a specific range of characteristics, but simply for whether, on the basis of our expectations, we are coping with the object. Coping need not be defined by any specific set of traits but rather by an ongoing mastery which Merleau-Ponty calls maximum grasp. What counts as maximum grasp varies with the goal of the agent and the resources of the situation. Thus it cannot be expressed in situation- free, purpose-free terms. To conclude: Pattern recognition is relatively easy for digital comput- ers if there are a few specific traits which define the pattern, but complex pattern recognition has proved intractable using these methods. Tran- scendental phenomenologists such as Husserl have pointed out that human beings recognize complex patterns by projecting a somewhat indeterminate whole which is progressively filled in by anticipated ex- periences. Existential phenomenologists such as Merleau-Ponty have related this ability to our active, organically interconnected body, set to respond to its environment in terms of a continual sense of its own functioning and goals. Since it turns out that pattern recognition is a bodily skill basic to all intelligent behavior, the question of whether artificial intelligence is pos- sible boils down to the question of whether there can be an artificial embodied agent. The question is philosophically interesting only if we restrict ourselves to asking if one can make such a robot by using a digital computer. (I assume there is no reason why, in principle, one could not The Role of the Body in intelligent Behavior / 163 construct an artificial embodied agent if one used components sufficiently like those which make up a human being.) A project to build such a digitally controlled robot is currently under way at M.I.T., and it is philosophically interesting to consider its pro- gram and its underlying assumptions. The project director, Minsky again, is modestly trying to make only a mechanical shoulder, arm, and hand, coordinated with a TV eye, but he proposes to make it use tools to construct things. The first simple task was to program a simplified robot arm to pick up blocks. This has indeed been accomplished and represents the early success one has learned to expect in the field. The problem which remains is, as usual, that of generalizing the present successful techniques. To bring a simple arm over to pick up a block requires locating the block in objective space, locating the arm in the same space, and then bringing the two together. This is already quite a feat. A mathematical description of the way an arm moves in objective space runs into surprising discontinuities. There are points which are contiguous in objective space which are far apart in reaching space. For example, to scratch our back we do not simply extend the position we use for scratching our ear. Living in our bodies we have built up a motor space, in which we sense these objectively contiguous points as far apart. We automatically reach for them in very different ways, and do not feel we have gone through the mathematics necessary to work out the opti- mal path for each specific case. For the programmer, however, who has to program the computer to calculate the movements of the mechanical arm in objective space, these discontinuities have so far proved an insur- mountable obstacle. The more flexible the arm the more degrees of freedom it has the more difficult and time consuming such calculations become. Rumor has it that an elaborate arm with six degrees of freedom, built by Minsky by 1965, has still not even been programmed to move, let alone pick up blocks or use tools. If one adds to this the fact that, in the case of any skill which takes place in real time (such as playing Ping-Pong), all calculations must be completed in real time (before the ball arrives), the outlook is not very promising. As Feigenbaum notes in his report on the current state of robot work: What Computers Can't Do / 164 Both the MIT and Stanford University groups have worked on programs for controlling a variety of arm-hand manipulators, from the very simple to the very complex, from the anthropomorphic variety to the very non-anthropomorphic. None of the more esoteric manipulators seems to have worked out very well, though there is no published documentation of successes, failures, and rea- sons. 27 In the light of these difficulties, what encourages researchers to devote their research facilities to such a project? Simply the conviction that since we are, as Minsky ingenuously puts it, "meat machines" and are able to play ping-pong, there is no reason in principle or in practice why a metal machine cannot do likewise. But before jumping to such a conclusion, the robot makers ought first to examine their underlying assumption that no essential difference exists between meat machines and metal ma- chines, between being embodied and controlling movable manipulators. How do human beings play ping-pong, or to make the matter simpler, how do human beings use tools? Heidegger, Merleau-Ponty, and Michael Polanyi have each devoted a great deal of thought to this question. Each discusses the important way that our experience of a tool we are using differs from our experience of an object. A blind man who runs his hand along the cane he uses to grope his way will be aware of its position and its objective characteristics such as weight, hardness, smoothness, and so forth. When he is using it, however, he is not aware of its objective position, its physical traits, nor of the varying pressure in the palm of his hand. Rather, the stick has become, like his body, a transparent access to the objects he touches with it. As Polanyi puts it: While we rely on a tool or a probe, these are not handled as external objects . . . they remain on our side . . . forming part of ourselves, the operating persons. We pour ourselves out into them and assimilate them as parts of our existence. We accept them existentially by dwelling in them. 28 In this way we are able to bring the probe into contact with an object in physical space without needing to be aware of the physical location of the probe. Merleau-Ponty notes that: The whole operation takes place in the domain of the phenomenal; it does not run through the objective world, and only the spectator, who lends his objective The Role of the Body in Intelligent Behavior / 165 representation of the living body to the active subject, can believe that . . . the hand moves in objective space. 29 But Merleau-Ponty admits that this ability seems "magical" from the point of view of science, so we should not be surprised to find that rather than have no explanation of what people are able to do, the computer scientist embraces the assumption that people are unconsciously running with incredible speed through the enormous calculation which would be involved in programming a computer to perform a similar task. However implausible, this view gains persuasiveness from the absence of an alter- native account. To make embodiment an acceptable alternative we will have to show how one could perform physical tasks without in any way appealing to the principles of physics or geometry. Consider the act of randomly waving my hand in the air. I am not trying to place my objective hand at an objective point in space. To perform this waving I need not take into account geometry, since I am not attempting any specific achieve- ment. Mow suppose that, in this random thrashing about, I happen to touch something, and that this satisfies a need to cope with things. (More about need in Chapter 9.) I can then repeat whatever I did this time in order to touch something without appealing to the laws necessary to describe my movement as a physical motion. I now have a way of bringing two objects together in objective space without appealing to any principle except: "Do that again." This is presumably the way skills are built up. The important thing about skills is that, although science requires that the skilled performance be described according to rules, these rules need in no way be involved in producing the performance. Human beings are further capable of remembering, refining, and reor- ganizing these somewhat indeterminate motor schemata. Piaget has amassed an enormous amount of evidence tracing the development of these motor skills, which he calls operations, and has come to a gestaltist conclusion: The specific nature of operations . . . depends on the fact that they never exist in a discontinuous state A single operation could not be an operation because the peculiarity of operations is that they form systems. Here we may well protest vigorously against logical atomism ... a grievous hindrance to the psychology of thought. 30 * What Computers Can't Do / 166 This same analysis helps dissipate the mistaken assumptions underly- ing early optimism about language translation. If human beings had to apply semantic and syntactic rules and to store and access an infinity of facts in order to understand a language, they would have as much trouble as machines. The native speaker, however, is not aware of having gener- ated multiple semantic ambiguities which he then resolved by appeal to facts any more than he is aware of having picked out complex patterns by their traits or of having gone through the calculations necessary to describe the way he brings his hand to a certain point in objective space. Perhaps language, too, is a skill acquired by innately guided thrashing around and is used in a nonrulelike way. Wittgenstein suggests this point when he notes, "In general we don't use language according to strict rules it hasn't been taught us by means of strict rules either." 31 Such a view is not behavioristic. Our ability to use language in a situation and in general the wholistic way the functional meaning organ- izes and structures the components of skilled acts cannot be accounted for in terms of the arbitrary association of neutral determinate elements any more than it can be analyzed in terms of their combination according to rules. If language is understood as a motor skill, we would then assimilate language and dwell in it the way we assimilate an instrument. As Polanyi puts it, To use language in speech, reading and writing, is to extend our bodily equip- ment and become intelligent human beings. We may say that when we learn to use language, or a probe, or a tool, and thus make ourselves aware of these things as we are of our body, we interiorise these things and make ourselves dwell in them.'-* Again, because we are embodied, the rules necessary to give an objective analysis of our competence need in no way be involved in our perfor- mance. The AI researcher and the transcendental phenomenologist share the assumption that there is only one way to deal with information: it must be made an object for a disembodied processor. For the transcendental The Role of the Body in Intelligent Behavior / 167 phenomenologist this assumption makes the organization of our intelli- gent behavior unintelligible. For the AI researcher it seems to justify the assumption that intelligent behavior can be produced by passively receiv- ing data and then running through the calculations necessary to describe the objective competence. But, as we have seen, being embodied creates a second possibility. The body contributes three functions not present, and not as yet conceived in digital computer programs: (1) the inner horizon, that is, the partially indeterminate, predelineated anticipation of partially indeterminate data (this does not mean the anticipation of some completely determinate alternatives, or the anticipation of com- pletely unspecified alternatives, which would be the only possible digital implementation); (2) the global character of this anticipation which determines the meaning of the details it assimilates and is determined by them; (3) the transferability of this anticipation from one sense modality and one organ of action to another. All these are included in the general human ability to acquire bodily skills. Thanks to this fundamental ability an embodied agent can dwell in the world in such a way as to avoid the infinite task of formalizing everything. This embodied sort of "information processing," in which the meaning of the whole is prior to the elements, would seem to be at work in the sort of complex pattern recognition such as speech recognition with which we began our discussion. It is also necessary, in order to account for our ability to recognize typicality, family resemblances, and similarity, where the objects recognized need have no traits in common at all. In all these cases individual features get their significance in terms of an underdetermined anticipation of the whole. If these global forms of pattern recognition are not open to the digital computer, which, lacking a body, cannot respond as a whole, but must build up its recognition starting with determinate details, then Oettinger is justified in concluding his speech recognition paper on a pessimistic note: "If indeed we have an ability to use a global context without recourse to formalization . . . then our optimistic discrete enumerative approach is doomed. . . ," 33 8 The Situation: Orderly Behavior Without Recourse to Rules In discussing problem solving and language translation we have come up against the threat of a regress of rules for determining relevance and significance. Likewise, in starting a learning process, something must be known before any rules can be taught or applied. In each case we have found that if there are no facts with fixed significance, only an appeal to the context can bring this regress to a halt. We must now turn directly to a description of the situation or context in order to give a fuller account of the unique way human beings are "in-the- world," and the special function this world serves in making orderly but nonrulelike behavior possible. To focus on this question it helps to bear in mind the opposing posi- tion. In discussing the epistemological assumption (Chapter 5) we saw that our philosophical tradition has come to assume that whatever is orderly can be formalized in terms of rules. This view has reached its most striking and dogmatic culmination in the conviction of AI workers that every form of intelligent behavior can be formalized. Minsky has even developed this dogma into a ridiculous but revealing theory of human free will. He is convinced that all regularities are rule governed. He therefore theorizes that our behavior is either completely arbitrary or it is regular and completely determined by rules. As he puts it: / 168 Orderly Behavior Without Recourse to Rules / 169 ". . . whenever a regularity is observed [in our behavior], its representa- tion is transferred to the deterministic rule region." 1 Otherwise our behavior is completely arbitrary and free. The possibility that our behav- ior might be regular but not rule governed never even enters his mind. We shall now try to show not only that human behavior can be regular without being governed by formalizable rules, but, further, that it has to be, because a total system of rules whose application to all possible eventualities is determined in advance makes no sense. In our earlier discussion of problem solving we restricted ourselves to formal problems in which the subject had to manipulate unambiguous symbols according to a given set of rules, and to other context-free problems such as analogy intelligence tests. But if CS is to provide a psychological theory and if AI programs are to count as intelligent they must extend mechanical information processing to all areas of human activity, even those areas in which people confront and solve open-structured problems in the course of their everyday lives. 2 * Open-structured problems, unlike games and tests, raise three sorts of difficulties: one must determine which facts are possibly relevant; which are actually relevant; and, among these, which are essential and which inessential. To begin with, in a given situation not all facts fall within the realm of possible relevancy. They do not even enter the situation. Thus, in the context of a game of chess, the weight of the pieces is irrelevant. It can never come into question, let alone be essential or inessential for deciding on a specific move. In general, deciding whether certain facts are relevant or irrelevant, essential or inessential, is not like taking blocks out of a pile and leaving others behind. What counts as essential depends on what counts as inessential and vice versa, and the distinction cannot be decided in advance, independently of some particular problem, or some particular stage of some particular game. Now, since facts are not relevant or irrelevant in a fixed way, but only in terms of human pur- poses, all facts are possibly relevant in some situation. Thus for example, if one is manufacturing chess sets, the weight is possibly relevant (al- though in most decisions involved in making and marketing chess sets, it will not be actually relevant, let alone essential). This situational character of relevance works both ways: In any particular situation an What Computers Can't Do / 170 indefinite number of facts are possibly relevant and an indefinitely large number are irrelevant. Since a computer is not in a situation, however, it must treat all facts as possibly relevant at all times. This leaves AI workers with a dilemma: they are faced either with storing and accessing an infinity of facts, or with having to exclude some possibly relevant facts from the computer's range of calculations. But even if one could restrict the universe for each particular problem to possibly relevant facts and so far this can only be done by the programmer, not the program the problem remains to determine what information is actually relevant. Even in a nonformal game like playing the horses which is much more systematic than everyday open-struc- tured problems an unlimited, indefinitely large number of facts remain as possibly relevant. In placing a bet we can usually restrict ourselves to such facts as the horse's age, jockey, past performance, and competition. Perhaps, if restricted to these facts from the racing form, the machine could do fairly well, possibly better than an average handicapper; but there are always other factors such as whether the horse is allergic to goldenrod or whether the jockey has just had a fight with the owner, which may in some cases be decisive. Human handicappers are no more omniscient than machines, but they are capable of recognizing the rele- vance of such facts if they come across them. The artificial intelligence approach to this human ability would have to be to give the machine knowledge about veterinary medicine, how people behave when they fight their employers, and so forth. But then the problem arises of sorting through this vast storehouse of data. To which the answer is that all this information would be properly coded and tagged in the machine memory so that the machine would just have to do a scan for "horse-race betting" and get out the relevant material. But not all relevant material would have been encoded with a reference to this particular use. As Charles Taylor has pointed out in an elaboration of this example: The jockey might not be good to bet on today because his mother died yester- day. But when we store the information that people often do less than their best just after their near relations die, we can't be expected to tag a connection with betting on horses. This information can be relevant to an infinite set of con- texts. Orderly Behavior Without Recourse to Rules / 777 The machine might select on the basis of the key concepts it was worrying about, horses, jockeys, jockey Smith, etc. and pick out all facts about these. But this too would give an absurdly wide scatter. Via jockey, man and horse, one would find oneself pulling out all facts about centaurs. The only way the machine could zero in on the relevant facts would be to take this broad class, or some other selected on such a broad swoop basis, and test to see whether each one had causal relevance to the outcome of the race, taking it into account if it had, and forgetting it if it hadn't. 3 * But if the machine were to examine explicitly each possibly relevant factor as a determinate bit of information in order to determine whether to consider or ignore it, it could never complete the calculations neces- sary to predict the outcome of a single race. If, on the other hand, the machine systematically excluded possibly relevant factors in order to complete its calculations, then it would sometimes be incapable of per- forming as well as an intelligent human to whom the same information was available. Even the appeal to a random element will not help here, since in order to take up a sample of excluded possibilities at random so that no possibility is in principle excluded, the machine would have to be pro- vided with an explicit list of all such other possibly relevant facts or a specific set of routines for exploring all classes of possibly relevant facts, so that no facts would be in principle inaccessible. This is just what could be done in a completely defined system such as chess, where a finite number of concepts determines totally and unequivocally the set of all possible combinations in the domain; but in the real world the list of such possibly relevant facts, or even classes of possibly relevant facts, would be indefinitely large ("infinite in a pregnant sense," to use Bar-Hillel's phrase). All the everyday problems whether in language translation, problem solving, or pattern recognition come back to these two basic problems: (1) how to restrict the class of possibly relevant facts while preserving generality, and (2) how to choose among possibly relevant facts those which are actually relevant. Even Minsky implicitly admits that no one knows how to cope with the amount of data which must be processed if one simply tries to store all facts: What Computers Can't Do / 112 At each moment in the course of thinking about a problem, one is involved with a large collection of statements, definitions, associations, and so on, and a net- work of goals. One has to deal not only with facts about objects, relations between objects, and the like, but also facts about facts, classes of facts, relations between such classes, etc. The heuristic programs that, as we shall see, so neatly demonstrate principles when applied to small models will not work efficiently when applied to large ones. Problems like looping, branching, measuring prog- ress, and generally keeping track of what is happening will come to require a disproportional part of the computation time. 4 Whatever it is that enables human beings to zero in on the relevant facts without definitively excluding others which might become relevant is so hard to describe that it has only recently become a clearly focused problem for philosophers. It has to do with the way man is at home in his world, has it comfortably wrapped around him, so to speak. Human beings are somehow already situated in such a way that what they need in order to cope with things is distributed around them where they need it, not packed away like a trunk full of objects, or even carefully indexed in a filing cabinet. This system of relations which makes it possible to discover objects when they are needed is our home or our world. To put this less metaphorically it is helpful to return to Charles Taylor's exten- sion of the horse-racing example. Much of a human being's knowledge of situations and their possibilities is know-how, that is, it cannot be exhaustively unpacked into a set of specific instructions or factual statements, but is a general capacity to generate appropri- ate actions and therefore, if necessary, the "instructions" underlying them. Usu- ally we think of this kind of indefinitely unpackable form of knowledge as bound up with the know-how which underlies our actions. But the same kind of knowl- edge underlies what we suffer, our "passions." Thus just as I have a general grasp on what it is to walk around, use my hands, drive a car, conduct a case in court (if I'm a lawyer), etc. So I have a general grasp on what it is to be threatened, to hear good news, to be jilted by my girl friend, to be made a fool of in public. Now the human handicapper has this general grasp of certain common human actions and passions. He has the sense of the race as a perilous enterprise which needs all the will and effort of jockey (and horse) to win. But included in this sense is the capacity to imagine or recognize an indefinite number of ways in which this will and effort could miscarry or be countered by fortune. These are not stored somewhere as separate facts in the mind or brain, they are not Orderly Behavior Without Recourse to Rules / 775 "unpacked"; they are just generatable from the general grasp of the situation. Of course, the general grasp of different men may differ in scope and exactitude. If the handicapper has ever ridden horses, then he has a much firmer grasp on the activity; he can sense a lot more finely what may go wrong. But even the city-bred gangster has some general grasp of what it is to fight and strain hard to win. But the artificial intelligence proponent may still want to protest that all this just represents an alternative method of "storage." Even if he admits that this method is not available to the machine, he might still ask how it solves the retrieval problem. How does the handicapper recognize just those odd factors which are relevant? The answer is that if we understand our grasp of the world as arising out of our dealing with it according to our different capacities, and our being touched by it according to our different concerns, then we can see that the problem of how a given concern or purpose conies to select the relevant features of our surroundings doesn't arise. For being concerned in a certain way or having a certain purpose is not something separate from our awareness of our situation; it just is being aware of this situation in a certain light, being aware of a situation with a certain structure. Thus being anxious for my own life because I have fallen among thugs is to sense the menace in that bulge in his pocket, to feel my vulnerability to his fist which might at any moment be swung at my face, and so on. 5 The human world, then, is prestructured in terms of human purposes and concerns in such a way that what counts as an object or is significant about an object already is a function of, or embodies, that concern. This cannot be matched by a computer, for a computer can only deal with already determinate objects, and in trying to simulate this field of con- cern, the programmer can only assign to the already determinate facts further determinate facts called values, which only complicates the re- trieval problem for the machine. In Being and Time Heidegger gives a description of the human world in which man is at home, on the model of a constellation of implements (Zuege), each referring to each other, to the whole workshop and ulti- mately to human purposes and goals. The directional signal on a car serves as an example of a "fact" which gets its whole meaning from its pragmatic context: The directional signal is an item of equipment which is ready-to-hand for the driver in his concern with driving, and not for him alone: those who are not travelling with him and they in particular also make use of it, either by giving What Computers Can't Do / 174 way on the proper side or by stopping. This sign is ready-to-hand within-the- world in the whole equipment-context of vehicles and traffic regulations. It is equipment for indicating, and as equipment, it is constituted by reference or assignment. 6 Wittgenstein too makes frequent references to human forms of life and concerns and to certain very general "facts of natural history" taken for granted in our use of language and in structuring our everyday activities facts, incidentally, of a very special kind which would presumably elude the programmer trying to program all of human knowledge. As Wittgenstein says, "The aspects of things that are most important for us are hidden because of their simplicity and familiarity. (One is unable to notice something because it is always before one's eyes.)" 7 Facts, more- over, which would be so pervasively connected with all other facts that even if they could be made explicit, they would be difficult if not impossi- ble to classify. The basic insight dominates these discussions that the situation is organized from the start in terms of human needs and pro- pensities which give the facts meaning, make the facts what they are, so that there is never a question of storing and sorting through an enormous list of meaningless, isolated data. Samuel Todes 8 * has described in detail the field-structure of experi- ence which is prior to the facts and implicitly determines their relevance and significance. He points out that the world is experienced as fields within fields. Bits or aspects of objects are not experienced as isolated facts but as nested in a series of contexts. And "in" has many different senses, none of them that of mere physical inclusion, which Minsky and McCarthy take as primary. Parts of objects are experienced as in objects which they comprise, objects are in places which they fill a place is situated in a local environment, which itself is in the horizon of possible situations in a human world. Data, then, are far from brute; aspects of objects are not given as directly in the world but as characterizing objects in places in a local environment in space and time in the world. We can and do zero in on significant content in the field of experience because this field is not neutral to us but is structured in terms of our interests and our capacity for getting at what is in it. Any object which we experience must appear in this field and therefore must appear in Orderly Behavior Without Recourse to Rules / 775 terms of our dominant interest at that moment, and as attainable by some variant of the activity which generated the field. Since we create the field in terms of our interests, only possibly relevant facts can appear. Relevance is thus already built in. In the horse race case, racing fits into a nested context of activities, games, sports, contests. To see an activity as a horse race is to organize it in terms of the intention to win. To return to Taylor's account: The handicapper is concerned to pick a winner. As a human being he has a sense of what is involved in the enterprise of winning, and his being concerned means that he is aware of a horse, jockey, etc., in a way in which dangers are salient. Hence he notices when he reads in the obituary columns that Smith's mother died yesterday (Smith being the jockey, and one he knows to be very susceptible), and for once he bets against the form. The machine would pick out Smith's mother's death, as a fact about Smith, along with all the others, such as that Smith's second cousin has been elected dogcatcher in some other city, etc., but will then have to do a check on the probable consequences of these different facts before it decides to take them into account or not in placing the bet. 9 Thus our present concerns and past know-how always already deter- mines what will be ignored, what will remain on the outer horizon of experience as possibly relevant, and what will be immediately taken into account as essential. Wittgenstein constantly suggests that the analysis of a situation into facts and rules (which is where the traditional philosopher and the computer expert think they must begin) is itself only meaningful in some context and for some purpose. Thus again the elements already reflect the goals and purposes for which they were carved out. When we try to find the ultimate context-free, purpose-free elements, as we must if we are going to find the ultimate bits to feed a machine bits that will be relevant to all possible tasks because chosen for none we are in effect trying to free the facts in our experience of just that pragmatic organiza- tion which makes it possible to use them flexibly in coping with everyday problems. Not that a computer model is ever really purpose-free; even a model in terms of information storage must somehow reflect the context, but such an analysis of context in terms of facts and rules is rigid and What Computers Can't Do / 176 restricting. To see this, let us grant that all the properties of objects (whatever that might mean) could be made explicit in a decision tree so that each node recorded whether the object has a certain situation- independent predicate or its converse. This sort of classification structure has been programmed by Edward Feigenbaum in his EPAM model. 10 * Such a discrimination net might, in principle, represent an exhaustive, explicit, apparently situation-free characterization of an object, or even of a situation, insofar as it was considered as an object. It thus seems to provide efficient information storage, while avoiding the field/object distinction. But something crucial is omitted in the description of such an information structure: the organization of the structure itself, which plays a crucial role in the informative storage. The information in the tree is differently stored and differently accessible depending on the order in which the discriminations are made. As William Wynn notes in a discus- sion of EPAM: EPAM's Classification process is ... too history-dependent and unadaptable, for the discrimination net can be grown only from the bottom down and cannot be reorganized from the top. Tests inserted in the net which later prove to be of little discriminatory power over a given stimulus set cannot be removed, nor can new tests be inserted in the upper portion of the net. Thus, once it is formed, EPAM's discrimination net is difficult to reorganize in the interest of greater retrieval efficiency. Any procedure that reorganizes the tests in the structure seriously impairs retrieval of many items held in the memory. 11 So the order of discriminations is crucial. But in the physical world all predicates have the same priority. Only the programmer's sense of the situation determines the order in the decision tree. Through the pro- grammer's judgment the distinction between the field and the objects in the field is introduced into the computerized model. The pragmatic context used by the programmer can indeed itself be characterized in a decision tree, but only in some order of discriminations which reflects a broader context. At each level information concerning this broader con- text is indeed embodied in the general structure of the tree, but at no particular node. At each level the situation is reflected in the pragmatic intuitions of the programmer governing the order of decisions; but this Orderly Behavior Without Recourse to Rules / 777 fixes the facts in one order based on a particular purpose, and inevitably introduces the lack of flexibility noted by Wynn. If, on the other hand, in the name of flexibility all pragmatic ordering could be eliminated so that an unstructured list of purified facts could be assimilated by machine facts about the sizes and shapes of objects in the physical world and even about their possible uses, as isolable functions then all these facts would have to be explicitly included or excluded in each calculation, and the computer would be overwhelmed by their infinity. This is not to deny that human beings sometimes take up isolated data and try to discover their significance by trying to fit them into a previ- ously accumulated store of information. Sherlock Holmes and all detec- tives do this as a profession; everyone does it when he is in a very unfamiliar situation. But even in these cases there must be some more general context in which we are at home. A Martian might have to proceed in a very unfamiliar context if he were on earth, but if he shared no human purposes his task of sorting out the relevant from the irrele- vant, essential from the inessential, would be as hopeless as that of the computer. We all know also what it is to store and use data according to rules in some restricted context. We do this, for example, when we play a game such as bridge, although even here a good bridge player stores data in terms of purpose and strategies and takes liberties with the heuristic rules. We also sometimes play out alternatives in our imagination to predict what will happen in the real game before us. But it is just because we know what it is to have to orient ourselves in a world in which we are not at home; or to follow rulelike operations like the heuristics for bidding in bridge; and how to model in our imagination events which have not yet taken place, that we know that we are not aware of doing this most of the time. The claim that we are nonetheless carrying on such operations unconsciously is either an empirical claim, for which there is no evidence, or an a priori claim based on the very assumption we are here calling into question. When we are at home in the world, the meaningful objects embedded What Computers Can't Do / 178 in their context of references among which we live are not a model of the world stored in our mind or brain; they are the world itself. This may seem plausible for the public world of general purposes, traffic regula- tions, and so forth. But what about my experience, one may ask; my private set of facts, surely that is in my mind? This seems plausible only because one is still confusing this human world with some sort of physi- cal universe. My personal plans and my memories are inscribed in the things around me just as are the public goals of men in general. My memories are stored in the familiar look of a chair or the threatening air of a street corner where I was once hurt. My plans and fears are already built into my experience of some objects as attractive and others as to be avoided. The "data" concerning social tasks and purposes which are built into the objects and spaces around me are overlaid with these personal "data" which are no less a part of my world. After all, personal threats and attractions are no more subjective than general human pur- poses. Now we can see why, even if the nervous system must be understood as a physical object a sort of analogue computer whose energy ex- change with the world must in principle be expressible as an input/out- put function, it begs the question and leads to confusion to suppose that on the information-processing level the human perceiver can be under- stood as an analogue computer having a precise I/O function reproduci- ble on a digital machine. The whole I/O model makes no sense here. There is no reason to suppose that the human world can be analyzed into independent elements, and even if it could, one would not know whether to consider these elements the input or the output of the human mind. If this idea is hard to accept, it is because this phenomenological account stands in opposition to our Cartesian tradition which thinks of the physical world as impinging on our mind which then organizes it according to its previous experience and innate ideas or rules. But even Descartes is not confused in the way contemporary psychologists and artificial intelligence researchers seem to be. He contends that the world which impinges on us is a world of pure physical motions, while the world "in the mind" is the world of objects, instruments, and so forth. Only the relation between these two worlds is unclear. Artificial intelligence Orderly Behavior Without Recourse to Rules / 779 theorists such as Minsky, however, have a cruder picture in which the world of implements does not even appear. As they see it, details of the everyday world snapshots, as it were, of tables, chairs, etc. are re- ceived by the mind. These fragments are then reassembled in terms of a model built of other facts the mind has stored up. The outer world, a mass of isolated facts, is interpreted in terms of the inner storehouse of other isolated, but well catalogued, facts which somehow was built up from earlier experiences of this fragmented world and the result is a further elaboration of this inner model. Nowhere do we find the familiar world of implements organized in terms of purposes. Minsky has elaborated this computer-Cartesianism into an attempt at philosophy. He begins by giving a plausible description of what is in fact the role of imagination: If a creature can answer a question about a hypothetical experiment without actually performing it, then it has demonstrated some knowledge about the world. For, his [sic] answer to the question must be an encoded description of the behavior (inside the creature) of some submachine or "model" responding to an encoded description of the world situation described by the question. 12 Minsky then, without explanation or justification, generalizes this plausi- ble description of the function of imagination to all perception and knowledge: Questions about things in the world are answered by making statements about the behavior of corresponding structures in one's model of the world. 13 He is thus led to introduce a formalized copy of the external world; as if besides the objects which solicit our action, we need an encyclopedia in which we can look up where we are and what we are doing: A man's model of the world has a distinctly bipartite structure: One part is concerned with matters of mechanical, geometrical, physical character, while the other is associated with things like goals, meanings, social matters, and the like. 14 If all knowledge requires a model we, of course, need a model of ourselves: When a man is asked a general question about his own nature, he will try to give a general description of his model of himself. 15 What Computers Can't Do / 180 And, of course, for this self-description to be complete we will need a description of our model of our model of ourselves, and so forth. Minsky thinks of this self-referential regress as the source of philosoph- ical confusions concerning mind, body, free will, and so on. He does not realize that his insistence on models has introduced the regress and that this difficulty is proof of the philosophical incoherence of his assumption that nothing is ever known directly but only in terms of models. In general the more one thinks about this picture the harder it is to understand. There seem to be two worlds, the outer data- and the inner data-structure, neither of which is ever experienced and neither of which is the physical universe or the world of implements we normally do experience. There seems to be no place for the physical universe or for our world of interrelated objects, but only for a library describing the universe and human world which, according to the theory, cannot exist. To dismiss this theory as incoherent is not to deny that physical energy bombards our physical organism and that the result is our experience of the world. It is simply to assert that the physical processing of the physical energy is not a psychological process, and does not take place in terms of sorting and storing human-sized facts about tables and chairs. Rather, the human world is the result of this energy processing and the human world does not need another mechanical repetition of the same process in order to be perceived and understood. This point is so simple and yet so hard to grasp for those brought up in the Cartesian tradition that it may be necessary to go over the ground once again, this time returning to a specific case of this confusion. As we have seen, Neisser begins his book Cognitive Psychology with an exposi- tion of what he calls "the central problem of cognition." There is certainly a real world of trees and people and cars and even books. . . . However, we have no direct, immediate access to the world, nor to any of its properties. 16 Here, as we have noted in Chapter 4, the damage is already done. There is indeed a world to which we have no immediate access. We do not directly perceive the world of atoms and electromagnetic waves (if it even Orderly Behavior Without Recourse to Rules / 181 makes sense to speak of perceiving them) but the world of cars and books is just the world we do directly experience. In Chapter 4 we saw that at this point, Neisser has recourse to an unjustified theory that we perceive "snapshots" or sense data. His further account only compounds the confusion: Physically, this page is an array of small mounds of ink, lying in certain positions on the more highly reflective surface of the paper. 17 But physically, what is there are atoms in motion, not paper and small mounds of ink. Paper and small mounds of ink are elements in the human world. Neisser, however, is trying to look at them in a special way, as if he were a savage, a Martian, or a computer, who didn't know what they were for. There is no reason to suppose that these strangely isolated objects are what men directly perceive (although one may per- haps approximate this experience in the very special detached attitude which comes over a cognitive psychologist sitting down to write a book). What we normally perceive is a printed page. Again Neisser's middle-world, which is neither the world of physics nor the human world, turns out to be an artifact. No man has ever seen such an eerie world; and no physicist has any place for it in his system. Once we postulate it, however, it follows inevitably that the human world will somehow have to be reconstructed out of these fragments. One-sided in their perspective, shifting radically several times each second, unique and novel at every moment, the proximal stimuli bear little resemblance to either the real object that gave rise to them or to the object of experience that the perceiver will construct as a result. 18 But this whole construction process is superfluous. It is described in terms which make sense only if we think of man as a computer receiving isolated facts from a world in which it has no purposes; programmed to use them, plus a lot of other meaningless data it has accumulated or been given, to make some sort of sense (whatever that might mean) out of what is going on around it. There is no reason to suppose that a normal human being has this problem, although some aphasics do. A normal person experiences the What Computers Can't Do / 182 objects of the world as already interrelated and full of meaning. There is no justification for the assumption that we first experience isolated facts, or snapshots of facts, or momentary views of snapshots of isolated facts, and then give them significance. The analytical superfluousness of such a process is what contemporary philosophers such as Heidegger and Wittgenstein are trying to point out. To put this in terms of Neisser's discussion as nearly as sense will allow, we would have to say: "The human world is the mind's model of the physical world." But then there is no point in saying it is "in the mind," and no point in inventing a third world between the physical and the human world which is an arbi- trarily impoverished version of the world in which we live, out of which this world has to be built up again. Oettinger, alone among computer experts, has seen that in the world of perception and language, where the linguist and artificial intelligence worker begins his analysis, a global meaning is always already present. What I want to suggest is not necessarily a novel suggestion; but it does seem to have been lost from sight, perhaps deservedly so, because, as I have pointed out, it doesn't tell one what to do next. What I suggest is that it almost seems as if the perception of meaning were primary and everything else a consequence of understanding meaning. 19 But Oettinger does not seem to see that if one simply looks for some new sort of process, by which this global meaning is "produced," thereby reversing the current misunderstanding, one is bound to find what seems a mystery or a dead end. When we try to turn this around and say, "Well now, here is this stream of sound coming at you or its equivalent on a printed page, and what is it that happens to your listening to me or in reading a printed page that enables you to react to the meaning of what I say?" we seem to hit a dead end at this point. 20 What Oettinger too fails to understand is that there is either a stream of sounds or there is meaningful discourse. The meaning is not produced from meaningless elements, be they marks or sounds. The stream of sounds is a problem for physics and neurophysiology, while on the level of meaningful discourse, the necessary energy processing has already taken place, and the result is a meaningful world for which no new Orderly Behavior Without Recourse to Rules / 183 theory of production is required nor can be consistently conceived. To avoid inventing problems and mysteries we must leave the physical world to the physicists and neurophysiologists, and return to our descrip- tion of the human world which we immediately perceive. The problem facing contemporary philosophers is to describe the context or situation in which human beings live, without importing prejudices from the history of philosophy or the current fascination with computer models. This brings us back to the problem of regularity and rules. Our context-guided activity in terms of which we constantly modify the relevance and significance of particular objects and facts is quite regular, but the regularity need not and cannot be completely rule gov- erned. As in the case of ambiguity tolerance, our activity is simply as rule governed as is necessary for the task at hand the task itself, of course, being no more precise than the rules. Wittgenstein, like Heidegger, sees the regulation of traffic as paradig- matic: The regulation of traffic in the streets permits and forbids certain actions on the part of drivers and pedestrians; but it does not attempt to guide the totality of their movements by prescription. And it would be senseless to talk of an 'ideal* ordering of traffic which would do that; in the first place we should have no idea what to imagine as this ideal. If someone wants to make traffic regulations stricter on some point or other, that does not mean that he wants to approximate to such an ideal. 21 This contextual regularity, never completely rule governed, but always as orderly as necessary, is so pervasive that it is easily overlooked. Once, however, it has been focused on as the background of problem solving, language use, and other intelligent behavior, it no longer seems necessary to suppose that all ordered behavior is rule governed. The rule-model only seems inevitable if one abstracts himself from the human situation as philosophers have been trying to do for two thousand years, and as computer experts must, given the context-free character of information processing in digital machines. The Situation as a Function of Human Needs We are at home in the world and can find our way about in it because it is our world produced by us as the context of our pragmatic activity. So far we have been describing this world or situation and how it enables us to zero in on significant objects in it. We have also suggested that this field of experience is structured in terms of our tasks. These are linked to goals, and these in turn correspond to the social and individual needs of those whose activity has produced the world. What does this tell us about the possibility of AI? If the data which are to be stored and accessed are normally organized in terms of specific goals, then it would seem that the large data base problem confronting AI could be solved if one just constructed a list of objectives and their priorities what computer workers dealing with decision-making pro- grams call a utility function and programmed it into the computer along with the facts. We have seen, however, that explicit objectives do not work, even for organizing simple problem-solving programs. The difficulties of simple means-ends analysis suggest that in order for the computer to solve even well-structured problems, it is not sufficient for the machine to have an objective and to measure its progress toward this preset end. Planning requires finding the essential operations, so "pragmatic considerations," for example, the relative importance of logical operations had to be The Situation as a Function of Human Needs / 185 surreptitiously supplied by the programmers themselves before the logic program could begin. We must now try to describe in more detail how this pragmatic structuring differs from means-ends analysis, ultimately asking, of course, whether this human capacity for purposive organiza- tion is in principle programmable on digital machines. The difference between human goals and machine ends or objectives has been noted by one scientist who has himself been working on pattern recognition. Satosi Watanabe describes this difference as follows: For man, an evaluation is made according to a system of values which is non- specific and quasi-emotive, while an evaluation for a robot could only be made according to a table or a specific criterion. . . . This difference is subtle but profound. [One might say] that a man has values while a machine has objectives. Certainly men too have objectives, but these are derived from a system of values and are not the final arbiter of his actions, as they would be for a robot. ... As soon as the objective is set the machine can pursue it just as the man can. Likewise human utilitarian behavior can be easily simulated by a machine if the quantitative utility and the probability of each alternative event is fixed and given to the machine. But a machine can never get at the source from which this utility is derived. 1 Watanabe claims that these values are essential to intelligent behavior. For one thing, as Watanabe points out, "there are infinitely many possi- ble hypotheses that are supported by experience. Limitation of these hypotheses to a smaller subset is often done by a vaguely conceived criterion, such as the principle of simplicity, or the principle of ele- gance." 2 More specifically, Watanabe argues that it can be demonstrated that any two objects have the same number of predicates in common. If this does not seem to us to be the case, it is because we consider certain predicates more important than others. This decision as to what is impor- tant depends on our system of values. 3 But why on our system of values and not on a list of objectives? How does what Watanabe calls a system of values differ from having a utility function? So far the only difference seems to be that values are vaguer. But throughout Watanabe's analysis there is no argument showing why these values are not just vague objectives which could be represented by a region on a quantitative scale. To understand this important difference, What Computers Can't Do / 186 which Watanabe has noted, but not explained, one must first abandon his way of posing the problem. To speak of values already gives away the game. For values are a product of the same philosophical tradition which has laid down the conceptual basis of artificial intelligence. Al- though talk of values is rather new in philosophy, it represents a final stage of objectification in which the pragmatic considerations which pervade experience and determine what counts as an object are conceived of as just further characteristics of independent objects, such as their hardness or color. A value is one more property that can be added to or subtracted from an object. Once he has adopted this terminology and the philosophical position it embodies, Watanabe is unable to explain how values differ from somewhat vague properties, and thus cannot explain why he feels they cannot be programmed. To understand the fundamen- tal difficulty Watanabe is trying to get at, we must be able to distinguish between objects, and the field or situation which makes our experience of objects possible. For what Watanabe misleadingly calls values belongs to the structure of the field of experience, not the objects in it. We have seen that experience itself is organized in terms of our tasks. Like the pattern of a chess game, the world is a field in which there are areas of attraction and repulsion, paths of accessibility, regions of activity and of repose. In our own perceptual world we are all master players. Objects are already located and recognized in a general way in terms of the characteristics of the field they are in before we zero in on them and concern ourselves with their details. It is only because our interests are not objects in our experience that they can play this fundamental role of organizing our experience into meaningful patterns or regions. Heidegger has described the way human concerns order experiences into places and regions: Equipment has its place or else it 'lies around': this must be distinguished in principle from just occurring at random in some spacial position. . . . The kind of place which is constituted by direction and remoteness (and closeness is only a mode of the latter) is already oriented towards a region and oriented within it Thus anything constantly ready-to-hand of which circumspective Being-in- the- World takes account beforehand has its place. The * where' of its readiness-to- hand is put to account as a matter for concern. . . .* The Situation as a Function of Human Needs / 187 Heidegger is also the first to have called attention to the way philoso- phy has from its inception been dedicated to trying to turn the concerns in terms of which we live into objects which we could contemplate and control. Socrates was dedicated to trying to make his and other people's commitments explicit so that they could be compared, evaluated, and justified. But it is a fundamental and strange characteristic of our lives that insofar as we turn our most personal concerns into objects, which we can study and choose, they no longer have a grip on us. They no longer organize a field of significant possibilities in terms of which we act but become just one more possibility we can choose or reject. Philoso- phers thus finally arrived at the nihilism of Nietzsche and Sartre in which personal concerns are thought of as a table of values which are arbitrarily chosen and can be equally arbitrarily abandoned or transvaluated. Ac- cording to Nietzsche, "The great man is necessarily a skeptic. . . . Freedom from any kind of conviction is part of the strength of his will." 5 * But what is missing in this picture besides a sense of being gripped by one's commitment? What difference does it make when one is trying to produce intelligent behavior that one's evaluations are based on a util- ity function instead of some ultimate concern? One difference, which Watanabe notes without being able to explain, is that a table of values must be specific, whereas human concerns only need to be made as specific as the situation demands. This flexibility is closely connected with the human ability to recognize the generic in terms of purposes, and to extend the use of language in a regular but nonrulelike way. Moreover, man's ultimate concern is not just to achieve some goal which is the end of a series; rather, interest in the goal is present at each moment structur- ing the whole of experience and guiding our activity as we constantly select what is relevant in terms of its significance to the situation at hand. 6 A machine table of objectives, on the other hand, has only an arbitrary relation to the alternatives before the machine, so that it must be explic- itly appealed to at predetermined intervals to evaluate the machine's progress and direct its next choice. Herbert Simon and Walter Reitman have seen that emotion and moti- vation play some role in intelligent behavior, but their way of simulating What Computers Can't Do / 188 this role is to write programs where "emotions" can interrupt the work on one problem to introduce extraneous factors or work on some other problem. 7 They do not seem to see that emotions and concerns accom- pany and guide our cognitive behavior. This is again a case of not being able to see what one would not know how to program. Heidegger tries to account for the pervasive concern organizing hu- man experience in terms of a basic human need to understand one's being. But this analysis remains very abstract. It accounts for significance in general but not for any specific goal or specific significance. Thus Heidegger in effect assimilates all human activity to creative problem solving or artistic creation where we do not fully know what our goal was until we have achieved it. For Heidegger there can be no list of specifications which the solution must fulfill. Still, our needs are determi- nate enough to give things specific meaning for us, and many of our goals are quite explicit. To understand this we require a more concrete phenomenological analysis of human needs. The philosophical and psychological tradition (with the exception of the pragmatists), however, has tried to ignore the role of these needs in intelligent behavior, and the computer model has reinforced this ten- dency. Thus N. S. Sutherland, Professor of Experimental Psychology at the University of Sussex, in an article "Machines and Men," writes: Survival and self maintenance are achieved by genetically building into the human brain a series of drives or goals. Some of the obvious ones are hunger, thirst, the sexual drive and avoidance of pain. All of these drives are parochial in the sense that one could imagine complex information processing systems exhibiting intelligent behavior but totally lacking them. 8 We have seen, however, that our concrete bodily needs directly or indirectly give us our sense of the task at hand, in terms of which our experience is structured as significant or insignificant. These needs have a very special structure, which, while more specific than Heidegger's account, does resemble artistic creation. When we experience a need we do not at first know what it is we need. We must search to discover what allays our restlessness or discomfort. This is not found by comparing various objects and activities with some objective, determinate criterion, The Situation as a Function of Human Needs / 189 but through what Todes calls our sense of gratification. This gratification is experienced as the discovery of what we needed all along, but it is a retroactive understanding and covers up the fact that we were unable to make our need determinate without first receiving that gratification. The original fulfillment of any need is, therefore, what Todes calls a creative discovery. 9 * Thus human beings do not begin with a genetic table of needs or values which they reveal to themselves as they go along. Nor, when they are authentic, do they arbitrarily adopt values which are imposed by their environment. Rather, in discovering what they need they make more specific a general need which was there all along but was not determinate. This is most obvious when dealing with less instinctual psychological needs. When a man falls in love he loves a particular woman, but it is not that particular woman he needed before he fell in love. However, after he is in love, that is after he has found that this particular relation- ship is gratifying, the need becomes specific as the need for that particu- lar woman, and the man has made a creative discovery about himself. He has become the sort of person that needs that specific relationship and must view himself as having lacked and needed this relationship all along. In such a creative discovery the world reveals a new order of significance which is neither simply discovered nor arbitrarily chosen. Soren Kierkegaard has a great deal to say about the way one's person- ality or self is redefined in such an experience, and how everything in a person's world gets a new level of meaning. Since such a change, by modifying a person's concerns, changes the whole field of interest in terms of which everything gets its significance, Kierkegaard speaks of these fundamental changes as changes in our sphere of existence. And because such a change cannot be predicted on the basis of our previous concerns, yet once it has taken place is so pervasive that we cannot imagine how it could have been otherwise, Kierkegaard speaks of a change of sphere of existence as a leap. 10 This same sort of change of world can take place on a conceptual level. Then it is called a conceptual revolution. Thomas Kuhn in his book The Structure of Scientific Revolutions has studied this sort of transforma- What Computers Can't Do / 790 tion. As he puts it: "Insofar as their only recourse to that world is through what they see and do, we may want to say that after a revolution scientists are responding to a different world." 11 The conceptual framework determines what counts as a fact. Thus during a revolution there are no facts to which scientists can appeal to decide which view is correct. 'The data themselves [have] changed. This is the [sense] in which we may want to say that after a revolution scientists work in a different world." 12 The idea that knowledge consists of a large store of neutral data, taken for granted by Minsky, is inade- quate to account for these moments of profound change. According to Kuhn, "there can be no scientifically or empirically neutral system of language or concepts." 13 What occurs during a scientific revolution is not fully reducible to a reinterpreta- tion of individual and stable data. In the first place the data are not unequivocally stable. A pendulum is not a falling stone, nor is oxygen dephlogisticated air. 14 This leads Kuhn to a rejection of the whole philosophical tradition which has culminated in the notion of reason as based on the storage and processing of "data." On the basis of his research Kuhn sees both the inadequacy of this tradition and why it nonetheless continues to seem self-evident. Are theories simply man-made interpretations of given data? The epistemologi- cal viewpoint that has most often guided Western philosophy for three centuries dictates an immediate and unequivocal, Yes! In the absence of a developed alternative, I find it impossible to relinquish entirely that viewpoint. Yet it no longer functions eifectively, and the attempts to make it do so through the introduction of a neutral language of observations now seem to me hopeless. 15 In suggesting an alternative view, or more exactly, in analyzing the way science actually proceeds so as to provide the elements of an alterna- tive view, Kuhn focuses on the importance of a paradigm, that is, a specific accepted example of scientific practice, in guiding research. Here, as in the case of family resemblance studied earlier, objects are under- stood not in terms of general rules but rather in terms of their relation to a specific concrete case whose traits or implications cannot be com- pletely formalized. The Situation as a Function of Human Needs / 191 [Scientists can] agree in their identification of a paradigm without agreeing on, or even attempting to produce, a full interpretation or rationalization of it. Lack of a standard interpretation or of an agreed reduction to rules will not prevent a paradigm from guiding research. . . . Indeed, the existence of a paradigm need not even imply that any full set of rules exist. 16 It is just this open-ended richness of paradigms which makes them important: Paradigms may be prior to, more binding, and more complete than any set of rules for research that could be unequivocally abstracted from them. 17 Without such paradigms scientists confront the world with the same bewilderment which we have suggested would necessarily confront an AI researcher trying to formalize the human form of life: In the absence of a paradigm ... all of the facts that could possibly pertain to the development of a given science are likely to seem equally relevant. 18 Indeed, without a paradigm it is not even clear what would count as a fact, since facts are produced in terms of a particular paradigm for interpreting experience. Thus finding a new paradigm is like a Kierke- gaardian leap: Just because it is a transition between incommensurables, the transition between competing paradigms cannot be made a step at a time, forced by logic and neutral experience. Like the gestalt switch, it must occur all at once (though not neces- sarily in an instant) or not at all. 19 Here it becomes clear that the idea of problem solving as simply storing and sorting through data with a specific end in view can never do justice to these fundamental conceptual changes, yet these changes determine the conceptual space in which problems can first be posed and in terms of which data get their pervasive character of relevance and significance, so that problems can be solved. The reigning conceptual framework implicitly guides r esearch just as the perceptual field guides our perception of objects. Finally, even more fundamental than these conceptual revolutions studied by Kuhn are cultural revolutions; for example, the beginning of Greek philosophy, as we have seen, set up a view of the nature of man What Computers Can't Do / 192 and rationality on which all subsequent conceptual revolutions have rung changes. Equally radically, with the beginning of Christianity a new kind of love became possible which was not possible in Greece; heroism became suspect as a sign of pride, and goodness came to consist in the sacrifices of saints. These cultural revolutions show us, as Pascal first pointed out, that there is no sharp boundary between nature and culture even instinctual needs can be modified and overridden in terms of paradigms thus there is no fixed nature of man. Man's nature is indeed so malleable that it may be on the point of changing again. If the computer paradigm becomes so strong that people begin to think of themselves as digital devices on the model of work in artificial intelligence, then, since for the reasons we have been rehearsing, machines cannot be like human beings, human beings may become progressively like machines. During the past two thousand years the importance of objectivity; the belief that actions are governed by fixed values; the notion that skills can be formalized; and in general that one can have a theory of practical activity, have gradually exerted their influence in psychology and in social science. People have begun to think of themselves as objects able to fit into the inflexible calculations of disembodied machines: machines for which the human form-of-life must be analyzed as a meaningless list of facts, rather than the flexible prera- tional basis of rationality. Our risk is not the advent of super-intelligent computers, but of subintelligent human beings. Conclusion This alternative conception of man and his ability to behave intelligently is really an analysis of the way man's skillful bodily activity as he works to satisfy his needs generates the human world. And it is this world which sets up the conditions under which specific facts become accessible to man in indefinite and open-ended ways, because these facts are origi- nally organized in terms of these needs. This enables us to see the fundamental difference between human and machine intelligence. Artifi- cial intelligence must begin at the level of objectivity and rationality where the facts have already been produced. It abstracts these facts 1 * from the situation in which they are organized and attempts to use the results to simulate intelligent behavior. But these facts taken out of context are an unwieldy mass of neutral data with which artificial intelli- gence workers have thus far been unable to cope. All programs so far "bog down inexorably as the information files grow." 2 No other data-processing techniques exist at present besides the ac- cumulation of facts, and once the traditional philosophical assumptions underlying work in artificial intelligence have been called into question there is no reason to suppose that digital data storage and retrieval techniques will ever be powerful enough to cope with the amount of data generated when we try to make explicit our knowledge of the world. Since the data about the world may well be infinite and the formalization / 793 What Computers Can't Do / 194 of our form-of-life may well be impossible, it would be more reasonable to suppose that digital storage techniques can never be up to the task. "Moreover, if this phenomenological description of human intelligence is correct, there are in principle reasons why artificial intelligence can never be completely realized. Besides the technological problem posed by storing a great number of bits of neutral data, there are in the last analysis no fixed facts, be they a million or ten million, as Minsky would like to believe. Since human beings produce facts, the facts themselves are changed by conceptual revolutions. Finally, if the philosopher or artificial intelligence researcher proposes to meet this objection by formalizing the human needs which generate this changing context, he is faced with the source of this same difficulty. Indeterminate needs and goals and the experience of gratification which guides their determination cannot be simulated on a digital machine whose only mode of existence is a series of determinate states. Yet, it is just because these needs are never completely determined for the individ- ual and for mankind as a whole that they are capable of being made more determinate, and human nature can be retroactively changed by individ- ual and cultural revolutions. CONCLUSION: THE SCOPE AND LIMITS OF ARTIFICIAL REASON 1O The Limits of Artificial Intelligence We are now in a position to draw together the various strands of our philosophical argument concerning the limits of artificial intelligence. The division of the field of artificial intelligence into two subfields, Cogni- tive Simulation (CS) and Artificial Intelligence (AI), has led to the treatment of two separate but interrelated questions: (1) Does a human being in "processing information" actually follow formal rules like a digital computer?, and (2) Can human behavior, no matter how gener- ated, be described in a formalism which can be manipulated by a digital machine? In discussing each of these questions we found, first, that the des- criptive or phenomenological evidence, considered apart from tradi- tional philosophical prejudices, suggests that nonprogrammable human capacities are involved in all forms of intelligent behavior. Moreover, we saw that no contrary empirical evidence stands up to methodological scrutiny. Thus, insofar as the question whether artificial intelligence is possible is an empirical question, the answer seems to be that further significant progress in Cognitive Simulation or in Artificial Intelligence is extremely unlikely. If in the face of these difficulties workers in artificial intelligence still wish to justify their optimism, the burden of proof is henceforth on them. They must show that despite the empirical difficulties artificial intelli- / 797 What Computers Can't Do / 198 gence must be possible. But the a priori case for artificial intelligence is even weaker here than the empirical one. The very arguments which are supposed to show that formalization must be possible turn out to be either incoherent or self-contradictory and show, on the contrary, that barring certain highly unlikely empirical assumptions which have been ruled out by common agreement, formalization is impossible. The a priori arguments for formalization thus turn into conditional in principle arguments against the possibility of CS and AI. Let us review these arguments in more detail. In discussing CS we found that in playing games such as chess, in solving complex problems, in recognizing similarities and family resemblances, and in using lan- guage metaphorically and in ways we feel to be odd or ungrammatical, human beings do not seem to themselves or to observers to be following strict rules. On the contrary, they seem to be using global perceptual organization, making pragmatic distinctions between essential and ines- sential operations, appealing to paradigm cases, and using a shared sense of the situation to get their meanings across. Of course, all this orderly but apparently nonrulelike activity might nonetheless be the result of unconsciously followed rules. But when one tries to understand this as a philosophical proposal that all behavior must be understood as following from a set of instructions, one finds a regress of rules for applying rules. This regress cannot be terminated by an appeal to ordinary facts for, according to the original claim, the facts must themselves always be recognized and interpreted by rule. One way to avoid this regress would be to claim that the ultimate data are inputs of physical energy and that such inputs can always be digital- ized and processed according to rule. This seems to be Fodor's view. The claim that these inputs are processed in a sequence of operations like a digital program is not unintelligible, but would, as Fodor admits, de- mand an incredibly complex formalism which no one has been able to discover or invent. In the absence of any empirical or a priori argument that such a formalism for processing physical inputs does or must exist, and given the empirical evidence that the brain functions like an ana- logue computer, there is no reason to suppose and every reason to doubt Conclusion / 799 that the processing of physical inputs in the human brain takes the form of a digital computer program. The only other way to avoid the regress of rules is to modify the thesis and claim that on the lowest level rules are automatically applied without instructions. But this leads to trouble in two ways: (1) Once the a priori thesis that all behavior must follow instructions is thus weakened, we might as well claim that skilled behavior need not be based on uncon- sciously followed instructions at any level, so the argument that in spite of the phenomenological evidence subjects must be following rules must be abandoned. (2) If one nonetheless insists that there must be an ultimate level of uninterpreted givens, and that the givens are neither physical inputs nor ordinary objects, one is left with the view that these givens must be impoverished bits of information about the human world. This gives us the notion of "stimulus information," the sense data or snapshots intro- duced by Neisser. But this a priori notion of stimulus information turns out to be incomprehensible. All that is given empirically are continuous physical inputs to the organism, on the one hand, and the world of ordinary objects given to the perceiving subject, on the other. No cogni- tive psychologist has succeeded in defining another sort of input between these two which would provide the ultimate bits of information to which the rules are to be applied. All accounts offered thus far turn out to be an incoherent mixture of physical description in terms of energy, and phenomenalist description in terms of crudely defined sense data. Thus the psychological claim that, appearances notwithstanding, in- telligent behavior is produced by following fixed formal rules like a digital computer is stuck with a regress of rules for applying rules. It can not extricate itself from this regress by appeal to a notion of physical input which it cannot use or stimulus input which it cannot define. Although there is no empirical evidence either from psychology or from the success of current work, AI workers, like workers in CS, are confident that a formalization of intelligent behavior must be possible. Their argument is never explicitly stated, but it seems to be based on an ontological assumption that the world can be analyzed into independent What Computers Can't Do / 200 logical elements and an epistemological assumption that our understand- ing of the world can then be reconstructed by combining these elements according to heuristic rules. The first claim is safe enough. Since he is not committed to describing human beings, the AI worker, unlike the cognitive psychologist, has no trouble identifying the ultimate bits to which the rules must be applied they are digitalized sound waves and the elements in the mosaic of a TV tube. These can be recognized without appeal to further rules. But the second claim, that these elements can be reassembled, when put forward as an a priori necessity, runs into a regress of higher and higher order rules, the converse of the regress of rules for applying rules faced by those in Cognitive Simulation. Since each of the logical elements is assumed to be independent of all the others, it has no significance until related to the other elements. But once these elements have been taken out of context and stripped of all significance it is not so easy to give it back. The significance to be given to each logical element depends on other logical elements, so that in order to be recognized as forming patterns and ultimately forming ob- jects and meaningful utterances each input must be related to other inputs by rules. But the elements are subject to several interpretations according to different rules and which rule to apply depends on the context. For a computer, however, the context itself can only be recog- nized according to a rule. Here again, too, this computer-dictated analysis conflicts with our experience. A phenomenological description of our experience of being- in-a-situation suggests that we are always already in a context or situa- tion which we carry over from the immediate past and update in terms of events that in the light of this past situation are seen to be significant. We never encounter meaningless bits in terms of which we have to identify contexts, but only facts which are already interpreted and which reciprocally define the situation we are in. Human experience is only intelligible when organized in terms of a situation in which relevance and significance are already given. This need for prior organization reappears in AI as the need for a hierarchy of contexts in which a higher or broader context is used to determine the relevance and significance of elements in a narrower or lower context. Conclusion / 201 Thus, for example, to note the relevance of and to disambiguate objects and utterances involving knives it is necessary to know whether one is in a domestic, medical, or combative context (among others). It is only in such contexts that the presence of knives becomes relevant and signifi- cant. Once the context is established, it can be used to interpret the objects or utterances so as to determine subcontexts. For example, the presence of knives in a domestic context will normally establish a nour- ishment subcontext where objects and utterances can be disambiguated as having to do with eating rather than aggression. But if each context can only be recognized in terms of features selected as relevant and interpreted in terms of a broader context, the AI worker is faced with a regress of contexts. As in the case of Cognitive Simulation, there might have been an empirical way out of the regress. Just as for CS the ultimate uninter- preted bits might have been digitalized physical inputs, here the ultimate context or set of contexts might have been recognizable in terms of certain patterns or objects which had a fixed significance and could be used to switch the program to the appropriate subcontext of objects or discourse. But again as in CS the evidence is against this empirical possibility. There do not seem to be any words or objects which are always relevant and always have the same significance the way the red spot of a female stickleback always means mating time to the male. There remains only one possible "solution." The computer program- mer can make up a hierarchy of contexts and general rules for how to organize them for the computer. He does this by appealing to his general sense of what is generally relevant and significant for a human being. In some situations, however, any fact may become important. To formalize this so that the computer could exhibit human flexibility, the program- mer would have to be able to make explicit all that he normally takes for granted in being a human being. However, once he tries to treat his own situation as if he were a computer looking at it from the outside, the computer programmer is himself faced with an infinity of meaning- less facts whose relevance and significance could only be determined in a broader context. Thus it turns out that a logical atomist ontology does not entail a What Computers Can't Do / 202 logical atomist epistemology. Even if the world is scanned into the computer in terms of logically independent bits, this does not mean that one can argue a priori that it can be reassembled. In fact the attempt to argue a priori that because the world can be resolved into bits it can be interpreted by formal rules ends up showing just the opposite. These considerations are supported by a general theory of human experience as being-already-in-a-situation in which the facts are always already interpreted. This theory also suggests that the ultimate situation in which human beings find themselves depends on their purposes, which are in turn a function of their body and their needs, and that these needs are not fixed once and for all but are interpreted and made determinate by acculturation and thus by changes in human self-interpretation. Thus in the last analysis we can understand why there are no facts with built-in significance and no fixed human forms of life which one could ever hope to program. This is not to say that children do not begin with certain fixed re- sponses in fact, if they did not, learning could never get started but rather that these responses are outgrown or overridden in the process of maturation. Thus no fixed responses remain in an adult human being which are not under the control of the significance of the situation. Could we then program computers to behave like children and boot- strap their way to intelligence? This question takes us beyond present psychological understanding and present computer techniques. In this book I have only been concerned to argue that the current attempt to program computers with fully formed Athene-like intelligence runs into empirical difficulties and fundamental conceptual inconsistencies. Whether a child computer could begin with situation-free responses and gradually learn depends on the role indeterminate needs and ability to respond to the global context play in learning. What work has been done on learning by Piaget, for example, suggests that the same forms of "information processing" are required for learning which are required for mature intelligent behavior, and that intelligence develops by "con- ceptual revolutions." This should not surprise us. Computers can only deal with facts, but man the source of facts is not a fact or set of facts, Conclusion / 203 but a being who creates himself and the world of facts in the process of living in the world. This human world with its recognizable objects is organized by human beings using their embodied capacities to satisfy their embodied needs. There is no reason to suppose that a world orga- nized in terms of these fundamental human capacities should be accessi- ble by any other means. The Future of Artificial Intelligence But these difficulties give us no idea of the future of artificial intelligence. Even if the attempt to program isolated intelligent activities always ultimately requires the programming of the whole mature human form of life, and even if an Athene-like digital computer is impossible in principle that is, even if mature human intelligence is organized in terms of a field which is reciprocally determined by the objects in it and capable of radical revision the question still remains to what extent workers in artificial intelligence can use their piecemeal techniques to approximate intelligent human behavior. In order to complete our analy- sis of the scope and limits of artificial reason we must now draw out the practical implications of the foregoing arguments. Before drawing our practical conclusions, however, it will be helpful to distinguish four areas of intelligent activity. We can then determine to what extent intelligent behavior in each area presupposes the four human forms of "information processing" we distinguished in Part I. This will enable us to account for what success has been attained and predict what further progress can be expected. One can distinguish four types of intelligent activity (see Table 1). We have seen that the first two types are amenable to digital computer simulation, while the third is only partially programmable and the fourth is totally intractable. Area I is where the S-R psychologists are most at home. It includes all forms of elementary associationistic behavior where meaning and context are irrelevant to the activity concerned. Rote learning of non- sense syllables is the most perfect example of such behavior so far pro- grammed, although any form of conditioned reflex would serve as well. What Computers Can't Do / 204 Table 1 CLASSIFICATION OF INTELLIGENT ACTIVITIES I. Associationistic II. Simple Formal I III. Complex Formal I IV. Nonfonnal Characteristics of Activity Irrelevance of mean- Meanings completely In principle, same as Dependent on ing and situation. explicit and situation II; in practice, in- meaning and independent. ternally situation- situation which dependent, indepen- are not explicit. dent of external situation. Innate or Learned by rule. Learned by rule and Learned by per- learned by practice. spicuous examples. repetition. Field of Activity (and Appropriate Procedure) Memory games, e.g., Computable or quasi- Uncomputable Ill-defined games, "Geography" (asso- computable games, games, e.g., chess or e.g., riddles (percep- ciation). e.g., trim or tic-tac- go (global intuition tive guess). toe (seek algorithm and detailed count- or count out). ing out). Maze problems Combinatorial prob- Complex combina- Open-structured (trial and error). lems (nonheuristic torial problems problems (insight). means/ends analysis). (planning and maze calculation). Word-by-word Proof of theorems Proof of theorems Translating a translation using mechanical where no mechanical natural language (mechanical proof procedures proof procedure (understanding in dictionary) . (seek algorithm). exists (intuition and context of use). calculation). Response to rigid Recognition of sim- Recognition of com- Recognition of patterns (innate ple rigid patterns, plex patterns hi varied and distorted releasers and classi- e.g., reading typed noise (search for patterns (recogni- cal conditioning). page (search for regularities). tion of generic or traits whose con- use of paradigm junction defines class case). membership). Kinds of Program Decision tree, list search, template. Algorithm. Search-pruning heuristics. None. Conclusion / 205 Also some games, such as the game sometimes called Geography (which simply consists of finding a country whose name begins with the last letter of the previously named country), belong in this area. In language translating, this is the level of the mechanical dictionary; in problem solving, that of pure trial-and-error search routines; in pattern recogni- tion, matching pattern against fixed templates. Area II is the domain of Pascal's esprit degeometrie the terrain most favorable for artificial intelligence. It encompasses the conceptual rather than the perceptual world. Problems are completely formalized and completely calculable. For this reason, it might best be called the area of the simple-formal. Here artificial intelligence is possible in principle and in fact. In Area II, natural language is replaced by a formal language, of which the best example is logic. Games have precise rules and can be calculated out completely, as in the case of nim or tic-tac-toe. Pattern recognition on this level takes place according to determinate types, which are defined by a list of traits characterizing the individuals which belong to the class in question. Problem solving takes the form of reducing the distance between means and ends by repeated application of formal rules. The formal systems in this area are simple enough to be manipulated by algorithms which require no search procedure at all (for example, Wang's logic program). Heuristics are not 01 j unnecessary here, they are a positive handicap, as the superiority of Wang's algorithmic logic program over Newell, Shaw, and Simon's heuristic logic program dem- onstrates. In this area, artificial intelligence has had its only unqualified successes. Area III, complex-formal systems, is the most difficult to define and has generated most of the misunderstandings and difficulties in the field. It contains behavior which is in principle reproducible but in fact intract- able. As the number of elements increases, the number of transforma- tions required grows exponentially with the number of elements involved. As used here, "complex-formal" includes those systems which in practice cannot be dealt with by exhaustive enumeration algorithms (chess, go, etc.), and thus require heuristic programs. 1 * Area IV might be called the area of nonformal behavior. This includes What Computers Can't Do / 206 all those everyday activities in our human world which are regular but not rule governed. The most striking example of this controlled impreci- sion is our disambiguation of natural languages. This area also includes games in which the rules are not definite, such as guessing riddles. Pattern recognition in this domain is based on recognition of the generic, or of the typical, by means of a paradigm case. Problems on this level are open-structured, requiring a determination of what is relevant and insight into which operations are essential, before the problem can be attacked. 2 * Techniques on this level are usually taught by generalizing from examples and are followed intuitively without appeal to rules. We might adopt Pascal's terminology and call Area IV the home of the esprit de finesse. Since in this area a sense of the global situation is necessary to avoid storing an infinity of facts, it is impossible in principle to use discrete techniques to reproduce directly adult behavior. Even to order the four as in Table 1 is misleadingly encouraging, since it suggests that Area IV differs from Area III simply by introducing a further level of complexity, whereas Area IV is of an entirely different order than Area III. Far from being more complex, it is really more primitive, being evolutionarily, ontogenetically, and phenomenologically prior to Areas II and III, just as natural language is prior to mathematics. The literature of artificial intelligence generally fails to distinguish these four areas. For example, Newell, Shaw, and Simon announce that their logic theorist "was devised to leam how it is possible to solve difficult problems such as proving mathematical theorems [II or III], discovering scientific laws from data [III and IV], playing chess [III], or understanding the meaning of English prose [IV]." 3 The assumption, made explicitly by Paul Armer of the RAND Corporation, that all intelligent behavior is of the same general type, has encouraged workers to generalize from success in the two promising areas to unfounded expectation of success in the other two. This confusion has two dangerous consequences. First there is the tendency, typified by Simon, to think that heuristics discovered in one field of intelligent activity, such as theorem proving, must tell us some- thing about the "information processing" in another area, such as the understanding of a natural language. Thus, certain simple forms of infor- Conclusion / 207 mation processing applicable to Areas I and II are imposed on Area IV, while the unique form of "information processing" in this area, namely that "data" are not being "processed" at all, is overlooked. The result is that the same problem of exponential growth that causes trouble when the techniques of Areas I and II are extended to Area III shows up in attempts to reproduce the behavior characteristic of Area IV. 4 * Second, there is the converse danger. The success of artificial intelli- gence in Area II depends upon avoiding anything but discrete, determi- nate, situation-free operations. The fact that, like the simple systems in Area II, the complex systems in Area III are formalizable leads the simulator to suppose the activities in Area III can be reproduced on a digital computer. When the difference in degree between simple and complex systems turns out in practice, however, to be a difference in kind exponential growth becoming a serious problem the programmer, unaware of the differences between the two areas, tries to introduce procedures borrowed from the observation of how human beings per- form the activities in Area IV for example, position evaluation in chess, means-ends analysis in problem solving, semantic considerations in theo- rem proving into Area III. These procedures, however, when used by human beings depend upon one or more of the specifically human forms of "information processing" for human beings at least, the use of chess heuristics presupposes fringe consciousness of a field of strength and weakness; the introduction of means-ends analysis eventually requires planning and thus a distinction between essential and inessential opera- tions; semantic considerations require a sense of the context. The programmer confidently notes that Area III is in principle formal- izable just like Area II. He is not aware that in transplanting the tech- niques of Area IV into Area III he is introducing into the continuity between Areas II and III the discontinuity which exists between Areas III and IV and thus introducing all the difficulties confronting the for- malization of nonformal behavior. Thus the problems which in principle should only arise in trying to program the "ill-structured," that is, open-ended activities of daily life, arise in practice for complex-formal systems. Since what counts as relevant data in Area III is completely explicit, heuristics can work to some extent (as in Samuel's Checker What Computers Can't Do / 208 Program), but since Area IV is just that area of intelligent behavior in which the attempt to program digital computers to exhibit fully formed adult intelligence must fail, the unavoidable recourse in Area III to heuristics which presuppose the abilities of Area IV is bound, sooner or later, to run into difficulties. Just how far heuristic programming can go in Area III before it runs up against the need for fringe consciousness, ambiguity tolerance, essential/inessential discrimination, and so forth, is an empirical question. However, we have seen ample evidence of trouble in the failure to produce a chess champion, to prove any interesting theorems, to translate languages, and in the abandonment of GPS. Still there are some techniques for approximating some of the Area IV short-cuts necessary for progress in Area III, without presupposing the foregoing human forms of "information processing" which cannot be reproduced in any Athena-like program. To surmount present stagnation in Area III the following improved techniques seem to be required: 1. Since current computers, even primitive hand-eye coordinating ro- bots, do not have bodies in the sense described in Chapter 7, and since no one understands or has any idea how to program the global organiza- tion and indeterminacy which is characteristic of perception and embod- ied skills, the best that can be hoped for at this time is some sort of crude, wholistic, first-level processing, which approximates the human ability to zero in on a segment of a field of experience before beginning explicit rule-governed manipulation or counting out. This cannot mean adding still further explicit ways of picking out what area is worth exploring further. In chess programs, for example, it is beginning to be clear that adding more and more specific bits of chess knowledge to plausible move generators, finally bogs down in too many ad hoc subroutines. (Samuel thinks this is why there has been no further progress reported for the Greenblatt chess program. 5 ) What is needed is something which corre- sponds to the master's way of seeing the board as having promising and threatening areas. Just what such wholistic processing could be is hard to determine, given the discrete nature of all computer calculations. There seem to be two different claims in the air. When Minsky and Papert talk of finding Conclusion / 209 "global features," they seem to mean finding certain isolable, and deter- minate, features of a pattern (for example, certain angles of intersection of two lines) which allow the program to make reliable guesses about the whole. This just introduces further heuristics and is not wholistic in any interesting sense. Neisser, however, in discussing the problem of seg- menting shapes for pattern recognition before analyzing them in detail makes a more ambitious proposal. Since the processes of focal attention cannot operate on the whole visual field simultaneously, they can come into play only after preliminary operations have already segregated the figural units involved. These preliminary operations are of great interest in their own right. They correspond in part to what the Gestalt psychologists called "autochthonous forces," and they produce what Hebb called "primitive unity." I will call them the preattentive processes to emphasize that they produce the objects which later mechanisms are to flesh out and interpret. The requirements of this task mean that the preattentive processes must be genuinely "global" and "wholistic." Each figure or object must be separated from the others in its entirety, as a potential framework for the subsequent and more detailed analyses of attention. 6 But Neisser is disappointing when it comes to explaining how this crude, first approximation is to be accomplished by a digital computer. He seems to have in mind simply cleaning-up heuristics which, as Neisser implicitly admits, only work where the patterns are already fairly clearly demarcated. "Very simple operations can separate units, provided they have continuous contours or empty spaces between them. Computer programs which follow lines or detect gaps, for example, are as easily written as those which fill holes and wipe out local irregularities." 7 But such techniques fail, for example, in the case of cursive script. Of course, it is hard to propose anything else. What is being asked for is a way of dealing with the field of experience before it has been broken up into determinate objects, but such preobjective experience is, by defi- nition, out of bounds for a digital computer. Computers must apply specific rules to determinate data; if the problem is one of first carving out the determinate data, the programmer is left with the problem of applying determinate rules to a blur. The best that can be hoped in trying to circumvent the techniques of What Computers Can't Do / 270 Area IV, therefore, may well be the sort of clever heuristics Minsky and Papert propose to enable a first-pass program to pick out certain specific features which will be useful in directing the program in filling in more details. But such a d hoc techniques risk becoming unmanageable and in any case can never provide the generality and flexibility of a partially determinate global response. 2. A second difficulty shows up in connection with representing the problem in a problem-solving system. It reflects the need for essential/ inessential discrimination. Feigenbaum, in discussing problems facing artificial intelligence research in the second decade, calls this problem "the most important though not the most immediately tractable." 8 He explains the problem as follows: In heuristic problem solving programs, the search for solutions within a problem space is conducted and controlled by heuristic rules. The representation that defines the problem space is the problem solver's "way of looking at" the problem and also specifies the form of solutions. Choosing a representation that is right for a problem can improve spectacularly the efficiency of the solution-finding process. The choice of problem representation is the job of the human program- mer and is a creative act. 9 This is the activity we called finding the deep structure or insight. Since current computers, even current primitive robots, do not have needs in the sense we have discussed in Chapter 9, and since no one has any idea how to program needs into a machine, there is no present hope of dispensing with this "creative act." The best that can be expected at this time is the development of programs with specific objectives which take an active part in organizing data rather than passively receiving them. Programmers have noticed that, in the analysis of complex scenes, it is useful to have the program formulate an hypothesis about what it would expect to find on the basis of data it already has, and look for that. This should not be confused with the way the human being organizes what counts as data in terms of his field of purposes. All that can be expected is fixed rules to apply to fixed data; that is, there will be a programmed set of alternatives, and the program can, on the basis of present data, select one of these alternatives as the most probable and_ look for further data on the basis of this prediction. Conclusion / 211 Thus, specific long-range objectives or a set of alternative long-range objectives might be built into game-playing and problem-solving pro- grams, so that in certain situations certain strategies would be tried by the computer (and predicted for the opponent). This technique, of course, would not remove the restriction that all these alternatives must be explicitly stored beforehand and explicitly consulted at certain points in the program, whereas human purposes implicitly organize and direct human activity moment by moment. Thus even with these break- throughs the computer could not exhibit the flexibility of a human being solving an open-structured problem (Area IV), but these techniques could help with complex-formal problems such as strategy in games and long-range planning in organizing means-ends analysis. 3. Since computers are not in a situation, and since no one under- stands how to begin to program primitive robots, even those which move around, to have a world, computer workers are faced with a final prob- lem: how to program a representation of the computer's environment. We have seen that the present attempt to store all the facts about the en- vironment in an internal model of the world runs up against the prob- lem of how to store and access this very large, perhaps infinite amount of data. This is sometimes called the large data base problem. Minsky's book, as we have seen, presents several ad hoc ways of trying to get around this problem, but so far none has proved to be generalizable. In spite of Minsky's claims to have made a first step in solving the problem, C. A. Rosen in discussing current robot projects after the work reported in Minsky's book acknowledges new techniques are still re- quired: We can foresee an ultimate capability of storing an encyclopedic quantity of facts about specific environments of interest, but new methods of organization are badly needed which permit both rapid search and logical deductions to be made efficiently. 10 In Feigenbaum's report, there is at last a recognition of the seriousness of this problem and even a suggestion of a different way to proceed. In discussing the mobile robot project at the Stanford Research Institute, Feigenbaum notes: What Computers Can't Do / 272 It is felt by the SRI group that the most unsatisfactory part of their simulation effort was the simulation of the environment. Yet, they say that 90% of the effort of the simulation team went into this part of the simulation. It turned out to be very difficult to reproduce in an internal representation for a computer the necessary richness of environment that would give rise to interesting behavior by the highly adaptive robot. 11 We have seen that this problem is avoided by human beings because their model of the world is the world itself. It is interesting to find work at SRI moving in this direction. It is easier and cheaper to build a hardware robot to extract what information it needs from the real world than to organize and store a useful model. Crudely put, the SRI group's argument is that the most economic and efficient store of information about the real world is the real world itself. 12 This attempt to get around the large data base problem by recalculat- ing much of the data when needed is an interesting idea, although how far it can go is not yet clear. It presupposes some solution to the wholistic problem discussed in 1 above, so that it can segment areas to be recog- nized. It also would require some way to distinguish essential from inessential facts. Most fundamentally, it is of course limited by having to treat the real world, whether stored in the robot memory or read off a TV screen, as a set of facts; whereas human beings organize the world in terms of their interests so that facts need be made explicit only insofar as they are relevant. What can we expect while waiting for the development and application of these improved techniques? Progress can evidently be expected in Area II. As Wang points out, "we are in possession of slaves which are . . . persistent plodders." 13 We can make good use of them in the area of simple-formal systems. Moreover, the protocols collected by Newell, Shaw, and Simon suggest that human beings sometimes operate like digital computers, within the context of more global processes. Since digital machines have symbol-manipulating powers superior to those of humans, they should, so far as possible, take over the digital aspects of human "information processing/' Conclusion / 273 To use computers in Areas III and IV we must couple their capacity for fast and accurate calculation with the short-cut processing made possible by fringe-consciousness, insight, and ambiguity tolerance. Leib- niz already claimed that a computer "could enhance the capabilities of the mind to a far greater extent than optical instruments strengthen the eyes." But microscopes and telescopes are useless without the selecting and interpreting eye itself. Thus a chess player who could call on a machine to count out alternatives once he had zeroed in on an interesting area would be a formidable opponent. Likewise, in problem solving, once the problem is structured and an attack planned, a machine could take over to work out the details (as in the case of machine-shop allocation or investment banking). A mechanical dictionary which could display meanings on a scope ranked as to their probable relevance would be useful in translation. In pattern recognition, machines are able to recog- nize certain complex patterns that the natural prominences in our experi- ence lead us to ignore. Bar-Hillel, Oettinger, and John Pierce have each proposed that work be done on systems which promote a symbiosis between computers and human beings. As Walter Rosenblith put it at a recent symposium, "Man and computer is capable of accomplishing things that neither of them can do alone." 14 Indeed, the first successful use of computers to augment rather than replace human intelligence has recently been reported. A theorem-prov- ing program called SAM (Semi-Automated Mathematics) has solved an open problem in lattice theory. According to its developers: Semi-automated mathematics is an approach to theorem-proving which seeks to combine automatic logic routines with ordinary proof procedures in such a manner that the resulting procedure is both efficient and subject to human intervention in the form of control and guidance. Because it makes the math- ematician an essential factor in the quest to establish theorems, this approach is a departure from the usual theorem-proving attempts in which the computer unaided seeks to establish proofs. 13 One would expect the mathematician, with his sense of relevance, to assist the computer in zeroing in on an area worth counting out. And this is exactly what happens. What Computers Can't Do / 214 The user may intervene in the process of proof in a number of ways. His selection of the initial formulas is of course an important factor in determining the course AUTO-LOGIC will take. Overly large or ill-chosen sets of initial formulas tend to divert AUTO-LOGIC to the proving of trivial and uninteresting results so that it never gets to the interesting formulas. Provided with a good set of initial formulas, however, AUTO-LOGIC will produce useful and interesting results. As the user sees that AUTO-LOGIC is running out of useful ways in which to use the original formulas, he can halt the process and insert additional axioms or other material. He can also guide the process by deleting formulas which seem unimportant or distracting. This real-time interplay between man and machine has been found to be an exciting and rewarding mode of operation. 16 Instead of trying to make use of the special capacities of computers, however, workers in artificial intelligence blinded by their early suc- cess and hypnotized by the assumption that thinking is a continuum will settle for nothing short of unaided intelligence. Feigenbaum and Feldman's anthology opens with the baldest statement of this dubious principle: In terms of the continuum of intelligence suggested by Armer, the computer programs we have been able to construct are still at the low end. What is important is that we continue to strike out in the direction of the milestone that represents the capabilities of human intelligence. Is there any reason to suppose that we shall never get there? None whatever. Not a single piece of evidence, no logical argument, no proof or theorem has ever been advanced which demon- strates an insurmountable hurdle along the continuum. 17 Armer prudently suggests a boundary, but he is still optimistic: It is irrelevant whether or not there may exist some upper bound above which machines cannot go in this continuum. Even if such a boundary exists, there is no evidence that it is located close to the position occupied by today's machines. 18 Current difficulties, once they are interpreted independently of opti- mistic a priori assumptions, however, suggest that the areas of intelligent behavior are discontinuous and that the boundary is near. The stagnation of each of the specific efforts in artificial intelligence suggests that there can be no piecemeal breakthrough to fully formed adult intelligent behavior for any isolated kind of human performance. Game playing, language translation, problem solving, and pattern recognition, each Conclusion / 215 depends on specific forms of human "information processing," which are in turn based on the human way of being in the world. And this way of being-in-a-situation turns out to be unprogrammable in principle using presently conceivable techniques. Alchemists were so successful in distilling quicksilver from what seemed to be dirt that, after several hundred years of fruitless efforts to convert lead into gold, they still refused to believe that on the chemical level one cannot transmute metals. They did, however, produce as by-products ovens, retorts, crucibles, and so forth, just as computer workers, while failing to produce artificial intelligence, have developed assembly programs, debugging programs, program-editing programs, and so on, and the M.I.T. robot project has built a very elegant mechani- cal arm. To avoid the fate of the alchemists, it is time we asked where we stand. Now, before we invest more time and money on the information-process- ing level, we should ask whether the protocols of human subjects and the programs so far produced suggest that computer language is appropriate for analyzing human behavior: Is an exhaustive analysis of human reason into rule-governed operations on discrete, determinate, context-free ele- ments possible? Is an approximation to this goal of artificial reason even probable? The answer to both these questions appears to be, No. Does this mean that all the work and money put into artificial intelli- gence have been wasted? Not at all, if instead of trying to minimize our difficulties, we try to understand what they show. The success and subse- quent stagnation of Cognitive Simulation and of AI, plus the omnipres- ent problems of pattern recognition and natural language understanding and their surprising difficulty, should lead us to focus research on the four human forms of "information processing" which they reveal and the situational character of embodied human reason which underlies them all. These human abilities are not necessary in those areas of intelligent activity in which artificial intelligence has had its early success, but they are essential in just those areas of intelligent behavior in which artificial intelligence has experienced consistent failure. We can then view recent work in artificial intelligence as a crucial experiment disaffirming the What Computers Can't Do / 216 traditional assumption that human reason can be analyzed into rule- governed operations on situation-free discrete elements the most im- portant disconfirmation of this metaphysical demand that has ever been produced. This technique of turning our philosophical assumptions into technology until