Dartmouth Conference 1956
The Dartmouth conference of 1956 was a summer workshop on artificial intelligence held at Dartmouth College in Hanover, New Hampshire, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, and widely treated as the founding event of AI as a named field. The planning document that preceded it is a separate matter; this entry concerns what actually took place in the summer: who came, what was shown, what was argued about, and why a gathering that fell short of its own ambitions still became the field's origin story.
In reality it was a loose, long and uneven workshop, not a single conference with a program. It ran for roughly two months in the summer of 1956, beginning in June,, and attendance was fluid: a handful of people stayed for long stretches while others dropped in for days or a couple of weeks. The four organizers were the sponsors; Shannon, at Bell Labs, came only for part of the time, and the original plan of a sustained, coordinated attack by about ten people on a list of topics did not materialize. Attendees named in later accounts include Ray Solomonoff, Oliver Selfridge, Trenchard More, Arthur Samuel of IBM, and Allen Newell and Herbert Simon, who came from Carnegie Tech and RAND. McCarthy and others have recalled that participants largely pursued their own projects and that the group did not reach consensus on methods.
The most consequential presentation was the Logic Theorist, from Newell, Simon and their RAND colleague Cliff Shaw. The program searched for proofs of theorems in the propositional logic of Whitehead and Russell's Principia Mathematica, using heuristics to prune an otherwise explosive tree of possible deductions, and in its published tests it proved 38 of the first 52 theorems in the second chapter. It was the one working demonstration of machine reasoning at the meeting, and it carried a conceptual claim — that heuristic search over symbol structures is the essence of problem solving — that shaped the symbolic AI tradition for decades. Newell and Simon, who preferred to describe their work as complex information processing, were less committed to the new label than the organizers, and the encounter set up the symbolic-AI approach and a lasting intellectual partnership and rivalry with the McCarthy and Minsky wing.
What the workshop produced was less a set of results than a community and a name. The phrase artificial intelligence, which McCarthy had put into the proposal's title, stuck, even though competitors such as cybernetics, automata studies and complex information processing were available and some participants were lukewarm about it. Individual threads had long afterlives: Solomonoff's work on inductive inference grew into algorithmic probability, Samuel's checkers learner became a landmark of machine learning, and Rochester's IBM group had simulated Hebbian learning in neural networks. The participants went on to found or lead the main centers of the 1960s: McCarthy and Minsky started the MIT AI project in 1959, McCarthy later founded the Stanford AI Laboratory, and Newell and Simon built the Carnegie Mellon program.
Several popular pictures of the event are inaccurate. It was not a large conference with published proceedings and resolutions; no single paper emerged from it, and records of who attended on which dates are partial and rest on later recollections. Nor was it the moment machines first showed intelligence, since work by Turing, Shannon's chess paper, Samuel's checkers program and the early neural-network models predated it. The anniversary is nonetheless a reasonable birthday for the field as an institution, because it was the first time the people who would define the discipline were gathered under that name, and the 2006 AI@50 commemoration at Dartmouth made that reading official.
Its modern relevance is partly ironic. The workshop's hopes — that significant progress on learning, language and abstraction could be made in one summer by a small group — became a standard example of early optimism about timelines. The dominant methods of the original attendees, symbolic reasoning and search, were later eclipsed in public attention by statistical learning and neural networks that the same meeting had included as a minor thread. Yet the questions raised there, how machines might learn, use language, form abstractions and improve themselves, are the central topics of today's AI, and the event remains the field's customary starting date.
Read the Dartmouth conference of 1956 as a modest, disorderly summer meeting whose true product was identity: a name, a roster of founders and one convincing demonstration, from which an entire field dated itself.