Dartmouth Proposal
The Dartmouth proposal is the document, dated August 31, 1955, in which John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon asked the Rockefeller Foundation to fund a summer study of artificial intelligence. Titled "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," it is the earliest known use of the phrase as the name of a research field, and it is read today less as a funding request than as a founding manifesto, the written statement of what the new field was supposed to attempt. The meeting it financed is covered under its own entry; here the subject is the text.
The authors came from four institutions, which says something about the proposal's pitch. McCarthy was an assistant professor of mathematics at Dartmouth, Minsky a Harvard Junior Fellow in mathematics and neurology, Rochester the manager of information research at IBM's Poughkeepsie laboratory, and Shannon a mathematician at Bell Telephone Laboratories and the founder of information theory. The combination of university mathematics, industrial computing and communications theory signaled that the project was meant to be a serious technical endeavor. The proposal described a two-month study by ten people, a size that reflected a plan for concentrated, collaborative work rather than a public conference.
Its central sentence is the conjecture on which the study was to proceed. The text says that the study would go forward 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. The sentence is a working hypothesis, not a result, and it carries the whole program: it commits to a representational view of intelligence in which anything describable can be programmed, and it turns the long philosophical question of whether machines can think into an engineering project of finding the right descriptions. The proposal then said the study would try to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.
The seven numbered topics sketch the field's early map. They were automatic computers, how a computer can be programmed to use a language, neuron nets, the theory of the size of a calculation, self-improvement, abstractions, and randomness and creativity. Each anticipated a strand that later became its own discipline: language processing, neural-network learning, computational complexity theory, metalearning and self-improving systems, representation learning, and the role of stochastic search in generating new ideas. The proposal also included individual statements of interest, with Shannon on applying information theory to machines and brain models, Minsky on machines that build abstract models of their environment, Rochester on originality arising from guided randomness, and McCarthy on the relation of language to intelligence.
The money and the name were both contested ground. The authors requested funds from the Rockefeller Foundation for salaries, travel and support of participants over the summer, and the grant that came was smaller than the request, roughly half by common accounts, which helps explain why the actual workshop was smaller and more improvised than the plan. As for the term, McCarthy later said he chose artificial intelligence deliberately, partly to set the work apart from cybernetics and from Norbert Wiener's orbit, and the label caught on despite objections that it was vague or overclaimed. It also shaped expectations: the phrase asserted that intelligence could be manufactured, which was exactly the thesis in the conjecture.
Readers often take two lessons from it that are mistaken. One is that it predicted rapid success; in fact it spoke of a study to make a significant advance on some of the problems, not of solving them in a summer, although the tone was optimistic and the implied timescale proved wildly short. The other is that it was an academic paper for publication; it was an administrative grant document that later became famous because of what followed. Its conjecture remains the shared premise of symbolic and statistical AI alike, and arguments for and against it, from Searle's Chinese room to debates about learned representations, are in effect arguments about the sentence it contains.
Read the Dartmouth proposal as a statement of faith dressed as a grant application: one conjecture, seven topics, a request that was only half granted, and a name that proved more durable than any of its promises.