4 October 2026 · History · about 17 minutes

Ten Men, One Summer: The Dartmouth Workshop of 1956 and the Naming of Artificial Intelligence

A nine-page proposal, a cheque for half what was asked, and a summer in which the host lost his list of attendees. This is the story of how artificial intelligence got its name, and what the name quietly left out.

The Conjecture

The document that named artificial intelligence opens with a single bold sentence: "The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."¹ Nobody in 1955 had established this conjecture. Nobody has established it since.

The document was a grant proposal, dated 31 August 1955 and submitted to the Rockefeller Foundation two days later. It asked for $13,500, ten participants, and two months at Dartmouth College in Hanover, New Hampshire. It introduced, for the first time in print, the phrase "artificial intelligence." The summer workshop it funded was smaller, shorter, and less conclusive than seventy years of retelling have suggested. Its name, however, has outlasted everything.

The Proposal

The proposal opened with its conjecture and then listed seven research problems: automatic computers and how to program them; how to make a computer use language; neuron nets; a theory of the size of calculations; self-improvement; abstraction; and randomness and creativity.² The document did not say how anyone would know whether progress on any of these had been made. It said that significant advances were likely if a carefully chosen group worked together for a summer.

Of the four signatories, Shannon was by some distance the most famous scientist on the page: he had already founded information theory, was co-editing a volume with McCarthy, and had results in cryptography and the theory of Turing machines. Minsky, a Harvard Junior Fellow in mathematics and neurology, had built a machine for simulating learning by nerve nets and completed a Princeton PhD thesis entitled "Neural Nets and the Brain Model Problem." Rochester was Manager of Information Research at IBM, jointly responsible for the design of the IBM Type 701, who had also worked on simulation of nerve nets to test theories in neurophysiology. McCarthy, the youngest, was an assistant professor of mathematics at Dartmouth who had worked on Turing machines, the speed of computers, and the use of languages by machines.³ The contrast between Shannon's standing and McCarthy's juniority is worth holding in mind: the youngest and least established man in the room had picked the name and written the central conjecture.

The Shadow of Turing

Before getting to the name and the money, it is worth pausing on what the proposal did not cite. Turing's 1950 paper "Computing Machinery and Intelligence," published in Mind, had asked whether machines could think and proposed what he called the imitation game as a practical test. It had been in print for five years when the Dartmouth proposal was written. The proposal does not cite it. McCarthy later acknowledged Turing's priority obliquely, noting that Turing "first understood that programming computers was the main way to realize AI."⁴ One scholar of the period has noted that the Dartmouth proposal did follow Turing closely in its working definition of the problem: the aim was to make "a machine behave in ways that would be called intelligent if a human were so behaving," a formulation that tracks Turing's almost exactly.⁵ But Turing had died in June 1954, fourteen months before the proposal was drafted, and cybernetics rather than Turing's paper was the dominant way of thinking about intelligent machines at the time. The proposal's silence on the 1950 paper reflects that context; it was looking to establish a new discipline, not to annotate an existing one.

The Name

The phrase "artificial intelligence" had never appeared in print before this proposal. McCarthy coined it, and he was explicit, years later, about why. In a document available from his Stanford archives, he wrote:

… one of the reasons for inventing the term "artificial intelligence" was to escape association with "cybernetics." Its concentration on analog feedback seemed misguided, and I wished to avoid having either to accept Norbert Wiener as a guru or having to argue with him.⁶

Cybernetics was the obvious existing label for the kind of work he wanted to pursue. Wiener had defined it as the science of control and communication in the animal and the machine, and had written the book that defined the field. He was brilliant, famous, and, by multiple contemporary accounts, a difficult collaborator, prone to dominating conversations and advancing his views on fields remote from his own.

The name was also, as McCarthy put it, a way of nailing a flag to a mast. He had been disappointed that papers in Automata Studies, the volume he had co-edited with Shannon that appeared in 1956, dealt mostly with automata theory and not with making machines behave intelligently. He wanted to direct researchers toward machine intelligence, and a bold name would help.⁷

Ronald Kline, drawing on unpublished Rockefeller Foundation archives for his 2011 history of the three-way split between cybernetics, automata studies, and the new AI, argued that the Dartmouth project's self-presentation as a break from cybernetics was partly a strategic act of demarcation: a new name helped carve out new intellectual territory, new funding lines, and new departments, even though many participants at Dartmouth still described their own work using the older vocabulary.⁸ Kline developed this argument further in his 2015 monograph The Cybernetics Moment (Johns Hopkins University Press), which remains the fullest scholarly treatment of the split. The break was less clean than the name implied, as the summer itself would show.

The clearest institutional evidence of how thin the break was lies in the Macy Conferences on Cybernetics, a series of interdisciplinary gatherings held between 1946 and 1953 under the auspices of the Josiah Macy Jr. Foundation. Those conferences had brought together Wiener, von Neumann, McCulloch, Shannon, Ashby, and others, several of whom would appear at Dartmouth in 1956, to work on precisely the questions the Dartmouth proposal claimed as new territory.⁹ The Wikipedia article on the Dartmouth workshop notes that the project postdated meetings such as the 1951 Paris cybernetics conference and the Macy meetings. McCarthy, characteristically, had attended the 1948 Hixon Symposium on Cerebral Mechanisms in Behavior as a new Caltech graduate student, and, after enrolling at Princeton for his PhD, gone to see von Neumann at the Institute for Advanced Study with his ideas on intelligent automata.¹⁰ Dartmouth was not a break from that lineage; it was its rebranding.

The Four Men

McCarthy had arrived at Dartmouth only in 1954. He had made friends with Minsky as a fellow graduate student in mathematics at Princeton, and had also met von Neumann during his Princeton years, going to the Institute for Advanced Study to present his ideas on intelligent automata.¹¹ He had spent the summer of 1955 at Rochester's IBM department in Poughkeepsie, where the two of them drew Minsky and Shannon into the project.¹² Shannon was eleven years older than McCarthy and had met him during McCarthy's summer at Bell Labs in 1952. Their working relationship was the spine of the proposal.

The standard popular account of AI's founding, Pamela McCorduck's Machines Who Think (first published 1979, revised 2004), drew on extensive interviews with these participants and established the story of the Dartmouth summer that most subsequent writers have repeated. The evidence from Ray Solomonoff's notebooks and Trenchard More's notes, examined more recently by Grace Solomonoff, tells a somewhat different story, and it is on those primary documents that this essay chiefly relies.¹³

The Money

McCarthy had first approached the Rockefeller Foundation in February 1955. In June, he and Shannon met with Robert Morison, the Foundation's Director of Biological and Medical Research, to discuss what kind of proposal might attract funding. Morison was unsure whether money would be made available for such a visionary project.¹⁴

Morison's response, once the proposal arrived in September 1955, was the first in a long series of moments in the history of artificial intelligence in which the person holding the money proved to be the most careful reader in the room. He offered $7,500 for five weeks and wrote back that the general feeling at the Foundation was that "this new field of mathematical models for thought, though very challenging for the long run, is still difficult to grasp very clearly. This suggests a modest gamble for exploring a new approach, but there is a great deal of hesitancy about risking any very substantial amount at this stage."¹⁵

Half the money requested, for five of the eight weeks proposed. The Rockefeller Foundation, which still calls the project five weeks long, the length it paid for, had funded the birth of artificial intelligence for the price of a modest academic conference. The man holding the cheque was the only person in the exchange who priced the claim against the evidence.

Who Was Planned, Who Came

On 26 May 1956, McCarthy notified Morison of eleven planned attendees. For the full period: Minsky, Julian Bigelow, Donald MacKay, Ray Solomonoff, John Holland, and McCarthy himself. For four weeks: Shannon, Rochester, and Oliver Selfridge. For the first two weeks only: Allen Newell and Herbert Simon of Carnegie Tech.¹⁶

In the event, MacKay and Holland did not come. Rochester was too busy working on the IBM 704 and sent his student Trenchard More in his place for three weeks. Shannon attended for parts of the workshop. The actual population was much smaller and more fluid than the plan. According to the attendance records kept by Solomonoff and More, since McCarthy lost his own list, only three people stayed for the full eight weeks: McCarthy, Minsky, and Ray Solomonoff. On the days when More was present and keeping count, between three and eight people attended the daily sessions. During the entire fifth week, the number was four: Etter, McCarthy, More, and Solomonoff.¹⁷

The full cast is harder to establish than the textbook version suggests. McCarthy afterwards circulated a list of forty-seven names, mixing actual participants with visitors and people who had merely expressed interest in the subject.¹⁸ Ray Solomonoff's own handwritten notes list twenty actual attendees; Trenchard More's list, arranged roughly by degree of involvement, runs to thirty-two. Among those who came, in addition to the organisers and Solomonoff and More, were Julian Bigelow, W. Ross Ashby, Warren McCulloch, Abraham Robinson, John Nash, David Sayre, Arthur Samuel, Alex Bernstein, Herbert Simon, and Allen Newell.¹⁹ Bernard Widrow recalled, in an oral history interview published by MIT Press, that he drove up from MIT with a colleague and spent a week listening to what people were saying: it seems to him Minsky was there, and McCarthy was there, and a man from IBM named Rochester was there, doing some very early work on neural nets.²⁰

They had the entire top floor of the Dartmouth mathematics department to themselves. Most weekdays, someone would lead a discussion about their ideas, or a more general conversation would range across the problems. They tried to find common meanings. One afternoon, by a detail preserved in Trenchard More's notes, they looked up the word "heuristic" in the dictionary. More guided the AI@50 conference organisers back to the building in 2006. A photograph taken in 2014 showed the dictionary still there, on the same stand, still open at the same page.²¹

Newell, Simon, and the Logic Theorist

Newell and Simon were down for the first two weeks only. McCarthy's own retrospective account says they came for "a few days," and they arrived with the most concrete thing anyone brought to Hanover: the Logic Theorist, a program designed to prove mathematical theorems by searching the space of valid logical steps.²² Their collaborator at RAND, J.C. Shaw, was writing the code for the JOHNNIAC computer; when it first ran there is uncertain, as below. The program had been described in a RAND report dated 15 June 1956, issued just before the workshop opened, under the title The Logic Theory Machine: A Complex Information Processing System.²³

At Dartmouth, Newell and Simon proudly presented it. The reception was lukewarm.²⁴ Minsky explained his own casual response, in an interview with the historian Pamela McCorduck, by saying that he had already sketched out a heuristic search procedure for a geometry machine and been able to hand-simulate it in an hour or so.

The Logic Theorist eventually proved 38 of the first 52 theorems in chapter two of Whitehead and Russell's Principia Mathematica, and for theorem 2.53 (often cited in the secondary literature as 2.85) it found a proof shorter and more elegant than the one Russell and Whitehead had published.²⁵ Whether it was actually running on a machine during the workshop, as opposed to having been hand-simulated, is a question the contemporary documents leave genuinely open. The RAND report of July 1956 describes a specified, hand-simulable system with machine realisation still ahead of it; one scholarly account dates the first JOHNNIAC-printed proof to August 1956, which would place it within the workshop's eight weeks.²⁶

The standard claim that the Logic Theorist was the only working program in the room at Dartmouth is not quite right. Arthur Samuel had come to the workshop with a checkers-playing program he had been developing at IBM since 1952. His first learning version was completed in 1955 and had been demonstrated publicly on television on 24 February 1956, before the workshop opened.²⁷ Solomonoff's notes record Samuel's talk on 6 August, in which he discussed board positions, intermediate goals, and scoring methods. Samuel was producing a heuristic program that did run, predating the Logic Theorist as an operational piece of software.²⁸ Nobody in 1956 wrote "the only working program in the room"; that sentence arrived later, once the summer needed to have meant something.

Herbert Simon, in January 1956, before coming to Dartmouth, had assembled his wife, their three children, and some graduate students in a room and handed each of them an index card representing one component of the Logic Theorist, asking them to act out the program by hand. It was, he later wrote, "nature imitating art imitating nature."²⁹ The whole project rested on the conjecture that whatever humans did when they reasoned could be decomposed precisely enough to be allocated to components, whether those components were neurons, transistors, or Simon's children holding index cards.

McCarthy, for his part, thought list structures were a great idea but did not like the IPL programming language Simon and Newell had developed; he immediately considered using Fortran augmented by list-processing operations, a thought that led, two years later, to Lisp.³⁰

What the Summer Produced, and Did Not

McCarthy had hoped that the whole group would meet for the whole summer and work collectively on specific problems. It did not happen, for reasons he later enumerated: the Rockefeller Foundation gave only half the money requested, and the participants had their own research agendas and were not much deflected from them.³¹ On 7 August, McCarthy sounded out the group on working jointly on chess, possibly using the IBM 704. Solomonoff's notes record that Minsky was very much against it, Shannon interested but unenthusiastic about chess's relevance to the artificial intelligence problem, and Rochester not keen but willing to go along if nothing better turned up. No group project materialised. McCarthy missed the deadline for the summary paper and the contributions he had hoped to place in that September's IRE Symposium proceedings; only Newell and Simon's Logic Theorist paper and a paper by Rochester and colleagues on neural cell assemblies appeared there.³²

The workshop produced no collective report and no agreed research programme. What it produced, in the first instance, was a network.

Ray Solomonoff's Assessment

Ray Solomonoff was perhaps the least famous person in the room, then and now. Minsky had recommended him to the other organisers as a "best buy," noting that he incessantly filled notebooks with observations and would therefore be invaluable when it came to writing any report of the project.³³ Minsky was right about the notebooks, if not about the report.

Solomonoff stayed for all eight weeks and wrote down what people said. At the end, he recorded a summary in four terse lines:

Th. research project wasn't very suggestive. The main things of value:

1. wrote and got report reproduced (very important)
2. Met some interesting people in this field
3. Got idea of how poor most thought in this field is
4. Some ideas: a) Search problem may be imp. b) These guys may eventually invent a T.M. simply by working more and more interesting special problems.³⁴

This is not the verdict one normally reads about the summer of 1956. Nor was Solomonoff wrong that the thinking in the field was poor. Most of the participants were pursuing deductive, logic-based approaches. Solomonoff was the only one working entirely on induction using probabilistic measures, and he recorded a disagreement that would define the next thirty years of AI research: reasoning machines built on non-probabilistic methods, he noted, could "be seriously disturbed forever by one ill-conceived example," while a probabilistic machine would be "not too much disturbed by a single counter example."³⁵ He was describing brittleness, and he lost the argument. Deductive and symbolic methods set the agenda until the 1980s.

Solomonoff's contribution was not entirely invisible. In a 2011 roundtable discussion recorded by Grace Solomonoff, Minsky said that Ray's early ideas about symbolic inductive inference had led him to switch from working on neural nets to working on symbolic ideas about AI: neural nets learned by computing correlations between things, whereas Ray's idea was to consider the possible meanings of different rearrangements of symbols.³⁶ The irony is that neural networks came back decisively in the 1980s and again in the 2000s, and are now the dominant approach; Solomonoff's probabilistic instinct has been retrospectively vindicated, though not in the form he imagined.

W. Ross Ashby on a Monday Afternoon

On 23 July 1956, both More and Solomonoff took notes on a talk by the British cybernetician W. Ross Ashby, a trained psychiatrist. Ashby described his homeostat, built in 1948: four units, each with a magnetic needle. When the system was stable the needles stayed centred. If the system became unstable, a needle would swing out and hit a stop; the machine would then switch one or more of its uniselectors, reorganising its own input connections, and search the new field of settings until stability returned. The way you tested it was by disturbing it. That was the only method of evaluation available, because a homeostat sitting undisturbed in a stable state tells you nothing at all about what it can do.³⁷

Ashby also stressed a point about machine evaluation that the field did not absorb: a simple machine, examined psychologically rather than mechanically, appears to be extraordinary. When part of how it works is concealed from the observer, the behaviour of the machine seems remarkable. Concealment inflates apparent capability. The objections recorded by both More and Solomonoff were correct: the homeostat kept no memory of previous solutions, so it took as long to handle a disturbance a second time as the first; and the time required to reach stability grew exponentially with the number of units. Both were problems of capability, and both were noted as such. Nobody in the room, as far as the notes show, drew attention to the evaluation method itself as the interesting part.

Ashby was, of course, a cybernetician, the tradition McCarthy had coined a new name to escape. The people best placed to ask how the new machines would fail were precisely the ones whose intellectual tradition had been set aside. Kline's archival research shows that many Dartmouth participants still described their own work using cybernetics vocabulary during the summer itself, even as the new label was taking hold outside the room.³⁸ On this essay's reading, the break from cybernetics was a naming decision before it was an intellectual one, and Ashby's visit suggests what the naming decision may have cost.

What the Name Did

The workshop ended around 17 August 1956, when Solomonoff gave a final talk. McCarthy sent Solomonoff's paper to Morison at the Rockefeller Foundation in November; Morison replied that they were naturally pleased to find that the seminar had progressed far enough that some of the thoughts had been worked out so fully.³⁹ It was a polite letter. It did not mention a collective report, because there was none.

What endured was not a research programme but a label and a concept. The name "artificial intelligence" gave a new field a clear identity distinct from cybernetics, from automata theory, from complex information processing. Departments could take the name, and did. Funding lines could adopt it, and did. Within two years, McCarthy and Minsky had started the MIT Artificial Intelligence Project; McCarthy later founded the Stanford AI Laboratory in 1962 and won the Turing Award in 1971; Newell and Simon won the 1975 Turing Award for the work that had its public debut in Hanover.⁴⁰ Rochester's enthusiasm for Minsky's geometry-machine idea, taken back to IBM, set Herbert Gelernter working on a geometry theorem prover, with McCarthy as a consultant, and Gelernter presented a major paper on the result in 1959.⁴¹ None of this was the result of the summer's collective work. It was the result of those weeks of conversation, disagreement, and proximity.

Two figures who might have shaped the proceedings were absent. McCarthy noted that Alan Turing, who had died in 1954, had understood that programming computers was the main way to realise artificial intelligence, an understanding already implicit in the proposal's seven problems. And John von Neumann, by the summer of 1956, was already ill from the cancer that killed him in early 1957.⁴² Their absences left the field to be shaped by people who were, in 1956, still working out their ideas.

The name, in the end, did more work than the summer. McCarthy chose it partly to avoid Norbert Wiener, and partly to nail a flag to a mast, and partly because it was neutral enough to cover the approaches he wanted to gather under one heading: symbolic reasoning, neuron nets, language, self-improvement, creativity. It was a name broad enough to become a field. One reading of the Ashby episode, and it is this essay's interpretation rather than something the notes establish, is that the name that escaped cybernetics also left behind cybernetics' central question: not what a machine can be made to do, but what happens when something disturbs it. Ashby tested his homeostat by pushing it off balance and watching what it did next. The seven problems in the proposal say nothing about disturbance. They say nothing about evaluation at all.

The conjecture underneath the proposal has never been established and may not be establishable. But it has been productive for seventy years. McCarthy measured the distance between what he had hoped to accomplish and what was accomplished, and called it large. Solomonoff put it more plainly still: the main thing of value was meeting some interesting people in the field. The history of the summer is, in part, the history of how those two verdicts got reversed in the telling.