19 October 2026 · History · about 13 minutes

The Philosopher in the Think Tank: Hubert Dreyfus, RAND, and the Paper That Infuriated Artificial Intelligence

A RAND consultant was hired to review AI research and found it philosophically incoherent. The field never quite forgave him, and never quite proved him wrong.

In 1964, Hubert Dreyfus sat in the offices of the RAND Corporation in Santa Monica and read himself into a fury. He was a philosopher, a specialist in Heidegger and Merleau-Ponty, and he had been hired as a consultant to evaluate the AI programme run by Allen Newell and Herbert Simon. What he found was a field that had achieved enough early successes, theorem-proving, puzzle-solving, the rudiments of chess, to convince its practitioners that general machine intelligence was nearly in hand, and that the remaining problems were technical rather than fundamental. Dreyfus thought the remaining problems were not technical at all. He thought the entire enterprise was founded on unexamined philosophical assumptions about the nature of mind, and he wrote this down in a report that grew to ninety-eight pages before RAND published it, in December 1965, as Paper P-3244: Alchemy and Artificial Intelligence.¹ Simon called it garbage.²

Dreyfus had been building to this moment for several years. From 1960 he had taught philosophy at MIT, eventually as associate professor, at the same moment MIT's Artificial Intelligence Laboratory was expanding under the direction of Marvin Minsky.³ His students in Heidegger seminars kept arriving with the same announcement: philosophy was finished. The computers down the hall were already solving the problems that Plato and Kant had only argued about. Dreyfus, who had spent a decade immersed in the phenomenology of Husserl and Merleau-Ponty, and who had interviewed Heidegger briefly at Freiburg on a Harvard Sheldon traveling fellowship in 1953–54, a meeting he reportedly found disappointing, was not persuaded.⁴

His first publication on the subject was a half-page objection to a talk given by Herbert Simon in the spring of 1961.⁵ It attracted little attention. The RAND paper would be a different matter.

A Consultant Is Hired

Dreyfus's brother Stuart was then working as a computer specialist at the RAND Corporation, and via this contact RAND recruited Dreyfus as a philosophical consultant to evaluate their AI programme.⁶ RAND's Paul Armer had thought he was hiring an impartial critic. He was surprised when Dreyfus produced a paper intended to demolish the foundations of the field. Armer stated he had been unaware of Dreyfus's earlier objection to Simon. He delayed publication, but in the end concluded that just because a paper came to a conclusion you did not like was no reason not to publish it.⁷

The RAND paper series, a product of the corporation from 1948 to 2003, was intended to capture the scholarly contribution of individual authors to their professional fields; papers in the series were less formal than reports and did not require rigorous peer review.⁸ The document was, in form, one scholar's contribution to a debate, not an official RAND finding. What happened next shows how little that distinction protected it.

The title announced Dreyfus's central analogy plainly. Alchemists, he wrote, had been so successful in distilling quicksilver from what seemed to be dirt that, after several hundred years of fruitless effort to convert lead into gold, they still refused to believe that on the chemical level one cannot transmute metals. To avoid the fate of the alchemists, it was time to ask where the AI programme actually stood.⁹

The Argument

He identified four assumptions, at least one of which he deemed necessary for AI to succeed. The first was what he called the biological assumption: that the brain processes information in discrete operations by way of some biological equivalent of on/off switches. The second was the psychological assumption: that the mind works by performing discrete computations on formal symbols. The third was the epistemological assumption: that all knowledge can be formalized in explicit, context-independent rules. The fourth was the ontological assumption: that reality consists of independent facts that can be represented by independent symbols.¹⁰

The paper argued that the attempt to analyse intelligent behaviour in digital computer language systematically excluded three fundamental human forms of information processing: what Dreyfus called fringe consciousness, essence/accident discrimination, and ambiguity tolerance. Of four distinct types of intelligent activity he identified, only two did not presuppose these human forms of information processing and could therefore be programmed. Significant developments in the remaining two areas would await an entirely different sort of computer, the only existing prototype for which was the little-understood human brain.¹¹

The philosophical backing for all this came from the phenomenological tradition Dreyfus had spent his career studying. Merleau-Ponty and Heidegger had argued, against the Cartesian tradition, that human beings do not relate to the world primarily through explicit representations and conscious rule-following. We act from habit, from context, from a kind of bodily familiarity with situations that is never made fully explicit and that Dreyfus doubted could ever be made explicit. When a carpenter reaches for a hammer, in Dreyfus's reading of Heidegger, he does not first consult a mental database of objects and their properties. The hammer is encountered as a thing-with-which-to-hammer, its meaning inseparable from the carpenter's situation and purpose. The world, in this view, is not a collection of independent facts waiting to be symbolically represented. It is a structured field of significance that human beings inhabit before they begin to reason about it.¹²

Underpinning this was a distinction that Gilbert Ryle had drawn in The Concept of Mind (1949): the difference between knowing-that, propositional knowledge, the kind that can be written down as rules, and knowing-how, the practical ability that an expert exercises without being able to articulate it as a set of instructions. Ryle had used it against Cartesian dualism; Dreyfus turned it against AI.¹³ The field, he argued, had succeeded precisely where knowing-that applied: formal logic, theorem-proving, games with explicit rules. It had made no progress where knowing-how was required. And most of what humans do, most of the time, is knowing-how.

Simon, following the success of the General Problem Solver, had in 1957 predicted that within ten years a digital computer would be the world chess champion, would discover and prove an important new mathematical theorem, and that most theories in psychology would take the form of computer programs.¹⁴ The press had dutifully reported these predictions as the imminent arrival of machine intelligence. Dreyfus, in 1965, argued that predictions like these would not come true.¹⁵

The Reaction

The paper caused what the historian Pamela McCorduck would later call an uproar.¹⁶ The AI community's response was derisive and personal. Seymour Papert dismissed one third of the paper as gossip and claimed that every quotation was deliberately taken out of context. Herbert Simon accused Dreyfus of playing politics so that he could attach the prestigious RAND name to his ideas, saying: "what I resent about this was the RAND name attached to that garbage."¹⁷

At MIT, where Dreyfus still taught, the consequences were social as well as intellectual. He recalled that his colleagues working in AI dared not be seen having lunch with him.¹⁸ The one exception was Joseph Weizenbaum, the author of ELIZA, who was an outspoken critic of Dreyfus's positions and thought his colleagues' treatment of him unprofessional. Weizenbaum later recalled: "I became the only member of the AI community to be seen eating lunch with Dreyfus. And I deliberately made it plain that theirs was not the way to treat a human being."¹⁹

Edward Feigenbaum, who would later be known as the father of expert systems, captured the reaction of the field to Dreyfus's use of phenomenology in a sentence that has outlasted most of his technical arguments: "What does he offer us? Phenomenology! That ball of fluff. That cotton candy!"²⁰ Marvin Minsky said of Dreyfus and the other critiques coming from philosophy that they misunderstood and should be ignored.²¹

Weizenbaum would go on to publish his own critique of AI, Computer Power and Human Reason, in 1976. Where Dreyfus asked what computers cannot do, Weizenbaum asked what they should not be allowed to do: his argument was that even if AI were possible, certain decisions, those requiring compassion, wisdom, and human judgment, should never be delegated to machines.²² The two critiques converged on the overreach of the field but came from different angles: Dreyfus's objection was philosophical and concerned with capability; Weizenbaum's was ethical and concerned with propriety.

Dreyfus's report was the first detailed critique of AI to be published, and almost immediately occupied the centre of a heated debate among computer scientists internationally.²³

Then came the chess match.

The Match

One of the things Dreyfus had noted in the paper was the state of computer chess. In 1965, no chess program could play even amateur chess, while Simon's ten-year deadline for a computer world champion was only two years from expiry.²⁴ The New Yorker picked this up in a short item on 11 June 1966, reporting with wry amusement on the victory of a ten-year-old child over the leading chess program.²⁵

By 1967, Richard Greenblatt, a programmer at MIT's AI Lab, had developed Mac Hack VI, a chess program written in MIDAS macro assembly language and running on a DEC PDP-6, the first computer program to play chess in human tournament competitions and to be granted a chess rating.²⁶ Seymour Papert is credited with having arranged a chess match between Dreyfus and the program, though Greenblatt's own oral history credits Jerry Sussman with bringing Dreyfus to the lab.²⁷ Dreyfus accepted. Dreyfus, playing white, lost in thirty-seven moves.²⁸ Greenblatt later noted dryly that Dreyfus was not a particularly good chess player.²⁹

The ACM's Special Interest Group in Artificial Intelligence published a bulletin with the headline: "A Ten Year Old Can Beat the Machine — Dreyfus: But the Machine Can Beat Dreyfus."³⁰ Dreyfus complained in print that he had never said a computer would never play chess. Simon replied: "You should recognize that some of those who are bitten by your sharp-toothed prose are likely, in their human weakness, to bite back... may I be so bold as to suggest that you could well begin the cooling — a recovery of your sense of humor being a good first step."³¹

What Was Actually Right

In 1972, Dreyfus expanded the RAND paper into a book, What Computers Can't Do: A Critique of Artificial Reason, which caused a second uproar. When McCorduck asked AI researchers why they had not mounted a sustained response, the answer she received was: "He's too silly to take seriously."³²

Dreyfus did not stop there. In 1986 he and his brother Stuart published Mind Over Machine, which elaborated the positive side of his argument. Drawing on studies of how human beings actually acquire skills, they proposed a five-stage model running from novice to expert. Novices follow explicit rules: the learner driver checks the speedometer before changing gear. Experts do not. As Dreyfus and Dreyfus wrote in Mind Over Machine: "when things are proceeding normally, experts don't solve problems and don't make decisions; they do what normally works."³³ The expert chess player does not search through decision trees; she sees the position. Expert systems, which tried to capture expertise by interviewing practitioners and encoding their answers as rules, were therefore doomed to capture only novice-level performance. This was Dreyfus's constructive claim, and it proved durable. The skill-acquisition model was later adopted widely outside AI, in nursing education, military training, and cognitive science, though it also attracted empirical pushback: Fernand Gobet and Philippe Chassy argued that there is no evidence for the presence of discrete stages and that experts in fact often engage in deliberate, analytical calculation.³⁴

In 1992, Dreyfus published the substantially revised What Computers Still Can't Do, adding a lengthy new introduction that assessed connectionism and neural networks. In it he accepted that sub-symbolic methods could capture some of the tendencies and attitudes he had described as essential to intelligence, an important concession, but maintained that they still fell short of the embodied, situated understanding he had identified as the real problem.³⁵

The field had, in its own way, arrived at a related finding. AI researchers named it Moravec's paradox, after the roboticist Hans Moravec, who articulated the observation in 1988 that it is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers, but difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility. The paradox was noted independently by Rodney Brooks, among others.³⁶ The skills that seemed easy were hard; the skills that seemed hard were easy. This is a restatement, in engineering terms, of what Dreyfus had argued philosophically in 1965.

The field also ran into the frame problem. First identified in a technical paper by John McCarthy and Patrick Hayes in 1969, the frame problem asks how a formal system can represent what does not change when an action occurs, how, in other words, a robot that has just moved a cup can know that everything else in the world remains as it was, without checking every fact in its database one by one. As the Stanford Encyclopedia of Philosophy notes, the problem generalises into the question of how any system of explicit representations could ever capture context-sensitive relevance. Dreyfus put it plainly: "if each context can be recognized only in terms of features selected as relevant and interpreted in a broader context, the AI worker is faced with a regress of contexts."³⁷ This was exactly the barrier Dreyfus had described philosophically in 1965. The engineers discovered it for themselves four years later.

By the early 1990s the grand predictions had not come true. Fully intelligent machines had not appeared in the mid-1970s as forecast. Natural language understanding remained stubborn. Common-sense reasoning, dealing with the open-ended, context-laden situations that human beings handle without effort, proved to be exactly the problem Dreyfus had said it would be. Daniel Crevier, the historian of AI, wrote that time had proven the accuracy and perceptiveness of some of Dreyfus's comments, and added that had he formulated them less aggressively, constructive actions they suggested might have been taken much earlier.³⁸

Connectionism, the research programme that culminated in the deep learning systems of the 2010s and after, dispensed with explicit symbol manipulation and replaced it with statistical learning from large amounts of data. McCorduck, writing in 2004, observed that this progress came piecemeal and in response to specific problems, and owes nothing to Dreyfus.³⁹ The people who built the systems that eventually did some of what Dreyfus said could not be done were not, for the most part, reading Merleau-Ponty. But it does not follow that Dreyfus was wrong. He had identified real problems that the field eventually addressed, not because it read him, but because it ran into the walls he had pointed to.

A few in computing took Heidegger seriously all the same. Terry Winograd and Fernando Flores, in Understanding Computers and Cognition (1986), proposed building AI, and computer systems generally, on Heideggerian principles.⁴⁰

Dreyfus himself was not modest about this. In 2007 he said: "I figure I won and it's over — they've given up."⁴¹ But his 2007 paper in the journal Artificial Intelligence, "Why Heideggerian AI Failed and How Fixing It Would Require Making It More Heideggerian", told a more complicated story. While acknowledging efforts at reorienting research beyond symbol-based AI toward what he called Heideggerian approaches, he rejected Brooks's behaviour-based robots as responding only to fixed features of the environment rather than learning new relevance in a worldly context, and concluded that no existing programme had adequately addressed the background problem he had raised in 1965.⁴² The triumphal remark and the scholarly verdict were in tension.

The Problem With the Critic

Dreyfus was not wrong on the substance. He was, however, combative in a way that made it easy for his opponents to treat his manner as a substitute for argument. McCorduck wrote that his derisiveness had been so provoking that he had estranged anyone he might have enlightened, and that this was a pity.⁴³

He was also limited by the philosophy he was deploying. Dreyfus drew on Heidegger and Merleau-Ponty, thinkers who were, in 1965, almost entirely unknown to AI researchers. Those researchers based their understanding of the human mind on engineering principles and efficient problem-solving techniques related to management science. On a fundamental level, they spoke a different language.⁴⁴ If someone tells you your foundations are rotten using a technical vocabulary you have never encountered, the temptation is to decide the vocabulary, not the foundation, is the problem.

There is also a genuine objection that Peter Norvig and Stuart Russell articulated clearly in their textbook, relaying a point Turing had made earlier: just because one cannot imagine formal rules that govern human intelligence does not mean no such rules exist.⁴⁵ Dreyfus's argument from phenomenology established that explicit rule-following was not how human beings seemed to experience their own intelligent behaviour. It did not conclusively establish that no underlying formal process could replicate that behaviour's results. The question of whether intelligence requires embodiment, or whether sub-symbolic statistical learning can substitute for it, remains genuinely open. The large language models of the 2020s handle the open-ended, context-laden language tasks that Dreyfus said would be forever beyond machines, and they do so without embodiment. Whether this vindicates his concession to connectionism or refutes his deeper claim, this essay's view, not the sources', is one of the more interesting questions currently facing the philosophy of mind.

The Last Lesson

Dreyfus was not a prophet. He was a philosopher who read the technical literature carefully, noticed what the engineers had assumed without examination, and said so in prose sharp enough to make enemies. The machines of 2026 can talk at length about a hammer without ever having held one. Whether that is the understanding he meant is the question he left behind, and nobody has yet settled it.