The Dartmouth Workshop — Naming Artificial Intelligence

On 31 August 1955, John McCarthy wrote a funding proposal to the Rockefeller Foundation. It opened: "We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire."

It is the earliest documented appearance of the phrase 'artificial intelligence' in an official research document.

Why Dartmouth

McCarthy was, at the time, an assistant professor of mathematics at Dartmouth, in his late twenties. He chose Dartmouth because it was his own institution. Four names appear on the proposal, with the affiliations it states:

• John McCarthy (Dartmouth College)
• Marvin Minsky (Harvard University) — described in the document as a Harvard Junior Fellow in Mathematics and Neurology, with a Princeton PhD thesis titled "Neural Nets and the Brain Model Problem"; he would later found the MIT AI Lab
• Nathaniel Rochester (I.B.M. Corporation) — Manager of Information Research at Poughkeepsie, and, in the proposal's words, one of the two engineers "jointly responsible for the design of the IBM Type 701"
• Claude Shannon (Bell Telephone Laboratories) — founder of information theory

The work would proceed, the proposal said, "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."

What the Proposal Actually Proposed

Below that opening paragraph the document lists seven aspects of the artificial intelligence problem. Read seventy years later, the list is less a period piece than a table of contents.

1. Automatic Computers. Capacity is not the bottleneck: "the major obstacle is not lack of machine capacity, but our inability to write programs taking full advantage of what we have."
2. How Can a Computer be Programmed to Use a Language. "It may be speculated that a large part of human thought consists of manipulating words according to rules of reasoning and rules of conjecture." The entry then concedes: "This idea has never been very precisely formulated nor have examples been worked out."
3. Neuron Nets. "How can a set of (hypothetical) neurons be arranged so as to form concepts." Existing work is credited to Uttley, Rashevsky's group, Farley and Clark, Pitts and McCulloch, and to Minsky, Rochester and Holland.
4. Theory of the Size of a Calculation. Trying every answer in order is a valid method and a useless one, so a criterion for efficiency is needed — which requires a measure of the complexity of calculating devices, which requires a theory of the complexity of functions. This is computational complexity, proposed a decade before the field had a name.
5. Self-Improvement. "Probably a truly intelligent machine will carry out activities which may best be described as self-improvement."
6. Abstractions. Classify the kinds of abstraction, then describe machine methods of forming abstractions from sensory and other data.
7. Randomness and Creativity. The conjecture "that the difference between creative thinking and unimaginative competent thinking lies in the injection of a some randomness" — the slip is in the document — guided by intuition rather than free.

Each of the four signatories attached a statement of what he personally intended to work on, and the four are not variations on one theme. Shannon proposed reliable computation from unreliable elements, plus what he called "the matched environment - brain model approach to automata": grow the model and its environment together, starting simple. Minsky proposed a machine that pairs motor abstractions with sensory ones so that "the machine would tend to build up within itself an abstract model of the environment in which it is placed", then tests solutions internally before acting. Rochester proposed to attack originality through controlled randomness, framing the whole question as: "how can I make a machine which will exhibit originality in its solution of problems?" McCarthy proposed language, listing properties English has that no formal language then had — conciseness, universality, self-reference, and the fact that "In addition to rules of proof, English if completely formulated would have rules of conjecture." He wanted an artificial language a computer could be programmed to use "on problems requiring conjecture and self-reference." By 1958 he was designing LISP.

The Money

The proposal is also a budget, and the budget is itemised:

The $1,200 was "for each faculty level participant who is not being supported by his own organization" — Bell Labs and IBM were expected to pay their own people, so foundation money was for Dartmouth and Harvard.

The Rockefeller Foundation, by its own account, backed the proposal with $7,500 — a little over half of what was asked, and less than the eight salaries alone would have cost. The famous "2 month, 10 man" line describes a plan that was never fully financed, which is one concrete reason it is not a roster.

One further deliverable never arrived. The list of people McCarthy circulated afterwards was headed by the statement that they would receive copies of the report of the Dartmouth Summer Project. Annotating his own document in 1996, McCarthy inserted a note of five words: "There was no report."

Who Was Expected, and Who Came

On 26 May 1956 McCarthy wrote to Robert Morison at the foundation naming eleven planned attendees and the time each would give: Minsky, Julian Bigelow, D. M. MacKay, Ray Solomonoff, John Holland and McCarthy for the full period; Shannon, Rochester and Oliver Selfridge for four weeks; Allen Newell and Herbert Simon for the first two. MacKay and Holland did not come at all, and Trenchard More stood in for Rochester for three weeks. McCarthy's covering note proposed an agenda that did not survive contact with the group: "we will concentrate on a problem of devising a way of programming a calculator to form concepts and to form generalizations. This of course is subject to change when the group gets together."

Others who appeared during the workshop include Arthur Samuel of IBM, author of an early learning program for checkers, and Tom Etter. There is no definitive attendee list. McCarthy lost his, and the list of 47 "People Interested in the Artificial Intelligence Problem" he circulated afterwards includes people who did not attend and omits some who did.

The Logic Theorist Was Still on Paper

The workshop ran from roughly 18 June to 17 August 1956. It is conventionally described as six weeks; Solomonoff's contemporaneous notes put it closer to eight. McCarthy, Minsky, and Solomonoff were the only three who stayed the whole time, and daily sessions drew somewhere between three and eight people.

The exhibit everyone remembers is the Logic Theorist, built at RAND by Allen Newell, Herbert Simon, and Cliff Shaw — and what was shown at Dartmouth is routinely overstated. In the summer of 1956 the program was not yet running on a machine. It was, in Newell and Simon's phrase, essentially specified: a paper design in the language later called IPL-I, whose behaviour they had checked by hand simulation, with family members and students acting out the parts of the program. Machine-executable proofs came later, from the IPL-II implementation on RAND's JOHNNIAC in late 1956 and early 1957.

What the design contained was the interesting part. LT worked in the sentential calculus using three rules of inference — substitution, replacement, and detachment — and four heuristic methods that generate subproblems instead of candidate proofs. The alternative, which the authors named the British Museum algorithm, was to grind out every derivable expression in order; by their own estimate that route "would take hundreds of thousands of years of computation to generate proofs for the theorems in Chapter 2" of Principia Mathematica. Heuristics buy speed by giving up guarantees, and the 1957 paper is candid about the price: the four methods are not a sufficient set, and the authors report finding a theorem, 2.13, that LT cannot prove at all.

From those 1957 runs comes the result usually attached to Dartmouth: 38 of the first 52 theorems of chapter two of Principia Mathematica, including a proof of theorem 2.85 more elegant than Whitehead and Russell's, which Simon showed to Russell himself. In March 2026 Jeff Shrager rebuilt an IPL-V interpreter in Common Lisp and ran the Logic Theorist again from Stefferud's 1963 transcription — 16 of 23 attempted theorems, and by his account the first execution of the original code in over half a century.

The workshop was not run as a single directed research project, and discussion ranged widely. No record survives of the participants converging on a shared theory.

A Name and a Network

Dartmouth is called the birth of artificial intelligence not for its technology but for its vocabulary and its network.

• Vocabulary. 'Artificial intelligence' became the field's official name. Before Dartmouth several names competed—'automata theory', 'thinking machines', 'cybernetics'.
• Network. The participants, and the networks branching out from them, were at the centre of AI research at MIT, Stanford, CMU, and IBM for the next two decades.

The first AI summer, from 1956 to the mid-1970s, ran on exactly this network. The papers they wrote, the students they trained, the ARPA (later DARPA) grants they secured—these built the discipline.

And underneath the first winter (late 1970s), the expert-system second summer (1980s), the second winter (early 1990s), and the present, still-running deep-learning summer (since 2012), the vocabulary chosen at Dartmouth has run beneath them all.