Dartmouth workshop

Also known as: Dartmouth Summer Research Project on Artificial Intelligence · Dartmouth conference

On this page 7

The Dartmouth workshop, formally the Dartmouth Summer Research Project on Artificial Intelligence, was an eight-week research gathering held at Dartmouth College in the summer of 1956 that gave artificial intelligence its name and is widely treated as the founding event of the field. Among AI fundamentals, it marks the point where scattered work on thinking machines became a self-identified research discipline with its own agenda, vocabulary and community.

The meeting is also called the Dartmouth conference, although it was closer to an extended, loosely organized workshop than a conference with sessions and published proceedings. Its lasting influence came less from specific results than from the research program it articulated and the people it connected.

How the Dartmouth proposal defined the field

The workshop began with a funding proposal dated August 31, 1955, written by John McCarthy of Dartmouth, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories. The document asked the Rockefeller Foundation to support a ten-person, two-month study of how machines might use language, form abstractions and improve themselves.

Its central claim was a working conjecture: that learning and every other feature of intelligence could, in principle, be described precisely enough for a machine to simulate it. This premise is the intellectual core of the event. It framed intelligence as a problem of formal description and computation, rather than as a question reserved for biology or philosophy.

The proposal listed seven research directions. The table below summarizes them and shows which later areas of the field each one anticipated.

Proposal topicCore questionLater research area
Automatic computersCan existing machines simulate higher mental functions?General AI programming
Language useHow can a computer manipulate words by rules?Natural language processing
Neuron netsHow can networks of neurons form concepts?Neural networks, connectionism
Size of a calculationHow efficiently can a problem be solved?Computational complexity, search
Self-improvementCan a machine improve its own behavior?Machine learning
AbstractionsHow can machines form abstractions from data?Representation learning, concept formation
Randomness and creativityCan controlled randomness yield original solutions?Stochastic search, heuristics
Research topics named in the 1955 proposal

Organizers, participants and the 1956 summer

The project ran for roughly eight weeks from mid-June to mid-August 1956 in Hanover, New Hampshire. The organizers requested about 13,500 US dollars and received a smaller grant of roughly 7,500 dollars, which limited stipends and travel. Attendance was fluid rather than fixed: some researchers stayed for weeks, while others visited for a few days.

Because no official roster or proceedings were produced, participant lists differ between historical accounts. The core figures are well documented, and the following table lists the most frequently cited contributors and their contemporary focus.

ParticipantAffiliation at the timeContribution or interest
John McCarthyDartmouth CollegeLead organizer; coined the field’s name
Marvin MinskyHarvard UniversityNeural nets, later symbolic reasoning
Claude ShannonBell LaboratoriesInformation theory, game-playing machines
Nathaniel RochesterIBMNeural network simulation on the IBM 701
Allen Newell and Herbert SimonRAND and Carnegie TechThe Logic Theorist program
Ray SolomonoffIndependent researcherInductive inference, probabilistic prediction
Oliver SelfridgeMIT Lincoln LaboratoryPattern recognition
Arthur SamuelIBMCheckers program that learned from play
Frequently cited participants and their focus in 1956

Why the name artificial intelligence mattered

McCarthy chose the term artificial intelligence for the proposal, and the workshop fixed it as the label of the discipline. The choice was partly strategic: it distinguished the new agenda from cybernetics, which focused on feedback and control, and from automata theory, which McCarthy considered too abstract. Not everyone agreed; Newell and Simon preferred “complex information processing” for years afterward.

The name carried consequences that persist today. It set ambitious expectations by inviting comparison with human minds, which later fed cycles of optimism and disappointment known as the AI winter. The table contrasts the Dartmouth framing with the neighboring traditions it competed against.

TraditionCentral ideaTypical methodRelation to Dartmouth
CyberneticsControl and feedback in animals and machinesAnalog circuits, servomechanismsDeliberately set apart
Automata theoryAbstract models of computing machinesMathematical proofsConsidered too narrow
Information processing psychologyCognition as symbol manipulationComputer simulation of reasoningOverlapping, different name
Dartmouth AI programIntelligence as precisely describable behaviorPrograms on digital computersDefined the new field
Dartmouth framing compared with neighboring traditions

What the workshop achieved and what it did not

The workshop produced no single breakthrough or shared theory. Participants largely pursued their own projects, and McCarthy later described the collaboration as weaker than he had hoped. The most concrete working result on display was the Logic Theorist, which Newell and Simon had built to prove theorems from Principia Mathematica, and it showed that symbolic AI could perform nontrivial reasoning.

Its real achievements were institutional and conceptual. Within a few years, the participants founded the major early AI laboratories and set research directions for two decades. The following table separates the short-term outputs from the longer-term consequences.

OutcomeTimeframeSignificance
Shared name and agenda1956Gave researchers a common identity
Logic Theorist demonstration1956Evidence that symbolic reasoning could be programmed
Lisp programming language1958McCarthy’s language for symbolic AI work
MIT AI project1959Founded by McCarthy and Minsky
Stanford AI Laboratory1963Founded by McCarthy after moving west
Carnegie AI research groupLate 1950sNewell and Simon’s work on problem solving
Short-term outputs and long-term consequences

Common misconceptions about the Dartmouth conference

Popular retellings often compress the story, and several details are routinely misstated. The points below correct the most frequent errors.

  • It was not the first AI research. Alan Turing’s 1950 paper on machine intelligence, Shannon’s chess analysis and early neural network models all came earlier.
  • It was not a formal conference. There were no fixed sessions, no peer review and no published proceedings.
  • It did not predict human-level AI within a summer. The proposal claimed only that a significant advance could be made on selected problems.
  • It did not favor symbolic methods exclusively. Neuron nets and randomness appeared in the agenda beside rule-based reasoning.
  • Its participants did not all agree. Disputes over naming and method began immediately and shaped later rivalries.

Understanding these nuances matters because the Dartmouth narrative is often used to support arguments about AI hype. A precise account shows an exploratory meeting whose optimism was real but whose stated claims were more modest than legend suggests.

Put simply

The Dartmouth workshop was a summer meeting in 1956 where a small group of scientists agreed that thinking could be studied by building computer programs. They gave this new area a name, artificial intelligence, and that name stuck.

Think of it as the founding meeting of a new club. Nobody solved the big problem that summer, but the members wrote down their goals, met each other and went home to start the first AI research labs. Much of the field’s later history traces back to the people who were there.

In practice

A university lecturer designing an introductory AI course uses the Dartmouth workshop as the anchor for the first week. Students read a short excerpt of the 1955 proposal and map each of its seven topics to a modern technique, such as linking self-improvement to gradient-based training and neuron nets to deep learning.

The exercise shows students that many current research questions are seventy years old, while the methods for answering them changed completely. It also prepares them for a discussion of why early predictions failed. By the end of the week, the class can explain the difference between the founding conjecture and the hype later attributed to it, which gives them a sober baseline for judging present-day claims about frontier AI.