There is a photograph taken at Dartmouth College in the summer of 1956 that has become one of the most reproduced images in the history of artificial intelligence. It shows a group of men — there are no women — standing in front of a building on the Dartmouth campus. They are dressed in the manner of American academics of the period: suits, ties, a general air of purposeful informality. Some are looking at the camera. Others are looking away, engaged in conversation with each other, distracted by whatever they were thinking about.

The photograph is not particularly dramatic. It does not look like a founding moment. It looks like a group of academics at a summer workshop.

But look at the names. John McCarthy. Marvin Minsky. Claude Shannon. Nathaniel Rochester. Allen Newell. Herbert Simon. These are the men who invented a field. They did not all know, in the summer of 1956, that this is what they were doing. They had come to work on a problem, not to found a discipline. The discipline happened anyway, as a byproduct of the work, as a consequence of the name that McCarthy had chosen, as the unintended result of two months of conversation in a small New Hampshire town.

This is the story of that summer. Not the sanitised founding myth — the clean narrative of visionaries gathered to launch a revolution. The real story, which is messier, more human, more interesting, and in some ways more important.

Why 1956 Was the Right Moment

Foundations are almost never accidents. The Dartmouth Conference happened in 1956 because 1956 was the right moment — the moment when the intellectual ingredients that had been accumulating for decades finally had a technological substrate capable of giving them physical form.

The theoretical foundations were in place. Turing had defined computation in 1936 and asked whether machines could think in 1950. Shannon had defined information in 1948. Von Neumann had designed the stored-program architecture in 1945. The McCulloch-Pitts model of the neuron had been published in 1943. The mathematical tools — computability theory, information theory, symbolic logic — were available.

The technological substrate was beginning to exist. The first generation of electronic computers had been built in the late 1940s. By 1956, there were working machines at universities and research laboratories across the United States. They were slow, unreliable, and expensive by modern standards — their memories were measured in kilobytes, their clock speeds in kilohertz, their reliability in hours between failures. But they were real. They could run programs. A person with the right idea and the right access could, for the first time in history, actually test whether a machine could exhibit something that resembled intelligence.

The institutional moment was also right. The post-war expansion of American science — driven by government investment in military research, by the Cold War competition with the Soviet Union, by a general cultural confidence in the power of science to solve problems — had created a landscape of research universities with resources and freedom that had not existed before. The Rockefeller Foundation and similar private foundations were funding interdisciplinary work that crossed the boundaries between established fields. DARPA was beginning to invest in speculative, high-risk research that private industry would not touch.

And there was a community — a small, interconnected, intensely productive community of people who had been thinking, from different angles and in different institutions, about related problems. Turing and the British computing tradition. Wiener and the cybernetics tradition. Shannon and the information theory tradition. Newell and Simon and the cognitive science tradition that was emerging from their work on the Logic Theorist. McCarthy and his colleagues working on mathematical logic and its relationship to computation. The community knew each other, read each other’s work, argued with each other at conferences. It was ready to cohere into something more organised than a network of individuals.

What was missing was a name. And a name, it turned out, was what John McCarthy was prepared to provide.

John McCarthy: The Man with the Most Important Idea

John McCarthy was twenty-eight years old in the summer of 1956. He had received his PhD from Princeton in 1951, spent time at Princeton and Stanford and Dartmouth, and was by 1956 an instructor in mathematics at Dartmouth College — a position that gave him access to the campus and the motivation to organise an event there.

He was, by the accounts of people who knew him, a man of formidable intellectual confidence. Not arrogance — he was genuinely interested in other people’s ideas and genuinely willing to change his mind when presented with good arguments. But confidence: a settled conviction that the problems he was working on were the right problems, that his approach to them was fundamentally correct, and that the people who disagreed with him were mistaken in ways he could explain.

This confidence expressed itself, in the Dartmouth proposal, in a choice that turned out to be one of the most consequential in the history of technology: the choice of the name “artificial intelligence.”

McCarthy was aware that other names were available. “Automata studies” was in use for work on the mathematical theory of computation. “Cybernetics” was Wiener’s term for the broader project of studying communication and control in machines and animals. “Complex information processing” was another possibility. “Machine intelligence” was how many British researchers described related work.

McCarthy chose “artificial intelligence” deliberately. He wanted a term that was direct about the goal — not the study of automata in general, or communication systems, or complex processing, but intelligence specifically, and intelligence as produced by human craft rather than by nature. The term was both precise and bold. It staked a claim: we are working on intelligence. That is our subject. That is what we intend to build.

The choice was not universally welcomed even among the conference participants. Norbert Wiener, who was not at Dartmouth but whose shadow fell across the gathering, had explicitly chosen “cybernetics” to avoid what he saw as the hubris implied by “artificial intelligence” — the implication that human intelligence was something that could simply be manufactured. McCarthy’s term seemed, to some, to claim too much.

But the term stuck. It stuck because it was clear, because it was specific, because it pointed directly at the central ambition rather than at a methodology or a related field. And because it stuck, it created a field: a community with a shared identity, a shared vocabulary, a shared set of problems, and eventually a shared institutional infrastructure of journals, conferences, departments, and research funding.

Names create communities. Communities do science. The name McCarthy chose in 1955, formalised at Dartmouth in 1956, made possible everything that followed.

The Proposal: A Document Worth Reading

The proposal that McCarthy submitted to the Rockefeller Foundation in 1955 — the document that officially launched the Dartmouth Conference and established the name “artificial intelligence” — is a remarkable piece of writing. It is short: a few pages, not long for a document of such historical consequence. It is direct: it states its objectives clearly and without unnecessary hedging. And it is, in retrospect, a mixture of genuine insight and spectacular overconfidence that captures both the strengths and the weaknesses of early AI with unusual clarity.

The proposal’s opening is its most famous passage: “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. 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.”

This conjecture — that intelligence is, at some level of description, a formal process that can be made explicit and mechanised — is the founding assumption of AI. It is not obviously true. Many philosophers, scientists, and ordinary people find it counterintuitive, even offensive — it seems to deny something important about the richness and mystery of human mental life. But it is the bet that the field has been making, in various forms, for seventy years. And the progress of the past few decades — the development of systems that can do things that once seemed to require specifically human intelligence — has provided substantial, if not conclusive, evidence that the bet was at least partially right.

The proposal then lists seven specific research areas that the conference would address:

  • Automatic computers
  • How can a computer be programmed to use a language?
  • Neuron nets
  • Theory of the size of a calculation
  • Self-improvement
  • Abstractions
  • Randomness and creativity

This list is both prescient and incomplete. Automatic computers — the general problem of computation — is the foundation. Language is central: it was the domain where the Turing Test had proposed that machine intelligence would be tested, and it remains the domain where the most impressive current AI systems operate. Neuron nets — what we would call neural networks — anticipated the approach that would ultimately prove most productive, though McCarthy himself was more sympathetic to symbolic approaches. Self-improvement — the ability of a system to improve its own performance — anticipates reinforcement learning and the meta-learning approaches of modern AI.

What the list does not mention is equally revealing. There is no mention of perception — of seeing or hearing — which turned out to be one of the hardest and most important problems in AI. There is no mention of common sense reasoning — the ability to navigate the everyday world using the vast, largely implicit knowledge that humans apply effortlessly — which turned out to be another of the hardest problems. There is no mention of learning from data — the statistical machine learning approach that would ultimately prove most productive. The proposal’s vision of AI was, inevitably, shaped by what the proposers could see from where they stood in 1955.

McCarthy’s optimism about timeline was, of course, the element most dramatically refuted by events. The belief that a summer’s work by ten carefully chosen researchers could make “significant advances” on most or all of these problems turned out to be wrong by decades. But the optimism was not irrational given what was known in 1955. The Logic Theorist had just proved theorems. Shannon’s information theory had provided a mathematical framework for thinking about communication. The first computers were operational. The problem seemed newly tractable.

The optimism was wrong about the timeline. The problem was right.

Hanover in June: The Setting and the People

The Dartmouth campus in summer is one of the more beautiful academic settings in New England — green hills, red brick buildings, a sense of remove from the urban intensity of the great research universities. The conference ran from June 18 to August 17, 1956, though the attendance was variable throughout — people came and went, stayed for days or weeks, returned after absences for other commitments.

The full roster of participants is worth examining in some detail, because the room that gathered at Dartmouth contained, in compressed form, a remarkable fraction of the intellectual capital that would build the field.

John McCarthy was the organiser and host. He was working, in the background of the conference, on what would become LISP — the programming language he would publish in 1958 that became the dominant language of AI research for thirty years. LISP was built around the manipulation of symbolic lists — a data structure that made it easy to represent and process the kinds of structured symbolic expressions that AI programs typically needed to work with. It was a language designed for the AI task, and it shaped how AI researchers thought about programming for decades.

Marvin Minsky was twenty-nine, fresh from his Harvard PhD, bristling with ideas and ambition. Minsky had built the SNARC — a neural network machine using forty vacuum tubes — as a student, but by 1956 he was moving away from neural networks toward symbolic approaches. He was interested in the question of how a machine could represent knowledge about the world — how it could build internal models that would allow it to reason about situations and plan actions. These interests would eventually lead him to frame theory, a rich but controversial approach to knowledge representation that he would develop over the following decades.

Claude Shannon appeared briefly. He was, by 1956, one of the most celebrated scientists in America, and his presence lent the conference enormous credibility. But he was not deeply engaged with the specific AI problems being worked on — his intellectual interests had moved in different directions since “A Mathematical Theory of Communication,” and he was characteristically reluctant to commit to positions he was not sure about. He participated in discussions, asked sharp questions, and left without having contributed a specific new result.

Nathaniel Rochester had been thinking seriously about whether a machine could develop internal representations through a process analogous to learning. Before Dartmouth, he had attempted to simulate a neural network on the IBM 701 — one of the first serious attempts to run a neural network simulation on a real computer. The simulation had not produced the results he hoped for, and this failure was on his mind at the conference.

Allen Newell and Herbert Simon arrived from Carnegie Tech with the Logic Theorist — the most impressive concrete result that anyone brought to the conference. They also brought, in preliminary form, ideas about the General Problem Solver — the more ambitious successor to the Logic Theorist that would attempt to implement a general problem-solving architecture applicable across domains. Their presentation of the Logic Theorist was the most dramatic moment of the conference: a working program that proved mathematical theorems, running on a real machine, demonstrated to an audience that had been trying to imagine what machine intelligence might look like.

Arthur Samuel was working on his checkers-playing program — a program that had already achieved a level of play that was surprising to people who understood how limited the computers of the era were. Samuel’s program used a technique he called rote learning: it remembered positions it had encountered before, along with how good or bad they had turned out to be, and used this memory to make better decisions in the future. This was a form of machine learning — the program improved with experience — and Samuel’s work would be enormously influential in subsequent AI research.

Ray Solomonoff was thinking about probability and prediction — about how a machine could learn to predict future events from past observations, and what mathematical framework could describe optimal inductive inference. His work, developed further in the years after Dartmouth, produced algorithmic information theory — a theory of the complexity and compressibility of data that connected Shannon’s information theory to Turing’s computability theory. It was deep, important, and in 1956 only beginning to take form.

Oliver Selfridge was interested in pattern recognition — in how a machine could learn to identify patterns in sensory data. He had developed a model of pattern recognition he called Pandemonium — an architecture in which multiple parallel processes competed to claim recognition of a pattern, with the most activated process winning. Pandemonium anticipated, in rough outline, the kind of hierarchical feature detection that convolutional neural networks would implement decades later.

Trenchard More completed the roster, working on mathematical logic and its connection to computing.

Together, these were the people at Dartmouth. Their backgrounds and interests were diverse, their approaches to the central problem were different, and their conversations over those two months — formal presentations, informal discussions, arguments over lunch and dinner — constituted the founding dialogue of the field.

What They Actually Did: The Days and Weeks

The Dartmouth Conference has been somewhat mythologised — described as a moment of unity and founding purpose that the historical record does not entirely support. The reality was more fragmented, more contested, and in some ways more interesting.

The conference was not organised around a formal programme of lectures or workshops. It was conceived as a research project — a collaborative effort to make actual progress on the problems listed in the proposal, not just to discuss them. McCarthy’s vision was of researchers working together on shared problems, producing results, not presenting pre-formed ideas to each other.

In practice, this vision did not fully materialise. The participants were strong-minded individuals with their own research agendas and their own approaches to the central problem. They were not inclined to subordinate their individual programmes to a collective enterprise, and the facilities for genuine collaboration — shared computing time on a single machine, regular joint working sessions — were limited. What happened instead was a series of individual presentations and bilateral conversations, with occasional larger group discussions when a particularly interesting result or argument demanded collective attention.

The most significant presentations were Newell and Simon’s Logic Theorist demonstrations and Shannon’s contributions on the theory of game-playing. Samuel showed his checkers program to people who found it impressive and somewhat unsettling — a machine that could genuinely learn from experience, improving its own performance without being explicitly reprogrammed, was something that most of the participants had not encountered before.

Minsky and McCarthy had extended conversations about the right architecture for AI — conversations that, while they did not produce immediate agreement, were shaping the thinking that both men would take back to their subsequent work. McCarthy’s emerging ideas about LISP — about the representation of programs and data as symbolic expressions that could be manipulated by other programs — were influenced by these conversations. Minsky’s thinking about knowledge representation and frame theory had its roots in the arguments he was having at Dartmouth about how a machine could model a domain well enough to reason about it effectively.

There were also arguments. The divide between the symbolic AI approach — reasoning from explicit representations, searching through logical structures — and the connectionist approach — learning from experience through the adjustment of network weights — was visible at Dartmouth from the beginning. McCarthy and Minsky were, in 1956, broadly aligned with the symbolic approach. Rochester’s neural network simulation had not worked well, which was discouraging for the connectionist side. But Solomonoff and Selfridge were thinking about learning and pattern recognition in ways that pointed toward statistical and connectionist approaches.

The arguments were not resolved at Dartmouth. They were not fully resolved for another sixty years.

The Evening Conversations: Where the Real Work Happened

The most productive moments of the Dartmouth Conference, by many accounts, were not the formal presentations but the informal conversations — the discussions over dinner, the debates that continued into the evening in dorm rooms and common areas, the bilateral exchanges between two researchers who found they were thinking about the same problem from different angles.

These conversations are less documented than the formal proceedings, but glimpses survive in memoirs, letters, and oral history interviews conducted decades later with surviving participants.

One recurring theme was the relationship between human intelligence and machine intelligence — the question of how closely a machine needed to resemble human cognitive processes in order to achieve intelligent behaviour. McCarthy’s position was broadly functionalist: what mattered was whether the machine could do what an intelligent being could do, not how it did it. If a chess program played chess brilliantly by searching game trees, it was playing chess brilliantly — it did not matter that it was not doing what a human chess player did internally.

Minsky was somewhat more concerned with the question of whether machine intelligence needed to be modelled on human intelligence — whether the architecture of human cognition provided a useful guide to the architecture of AI systems. His interest in neuroscience and his early work on neural networks reflected this concern. But he was also sceptical of naive analogies between brains and computers — he did not think that simply building systems that resembled neurons would automatically produce intelligent behaviour.

Shannon was more agnostic and more questioning. He asked sharp questions about what the participants actually meant by intelligence, what would count as evidence that a machine had achieved it, and what the relationship was between the specific capabilities being demonstrated and the broader goal of general machine intelligence. These were good questions. They were not always welcome.

The question of timelines came up repeatedly. How long would it take to build a genuinely intelligent machine? The estimates varied, but they were uniformly optimistic by standards that history would subsequently expose. Simon’s famous prediction — that within ten years a computer would be the world’s chess champion and would discover and prove an important new mathematical theorem — was made in this atmosphere, in the confidence that the Logic Theorist’s success was just the beginning of a rapid progression.

What is striking about these predictions, in retrospect, is not that they were wrong about the timelines. It is that the things they predicted eventually happened — but with delays of decades rather than years. A computer became the world’s chess champion in 1997, not 1967. AI systems proved important new mathematical theorems — though “important” and “new” are contested — within decades rather than within years. The vision of where AI was going was, broadly, correct. The pace at which it would get there was not.

The Intellectual Legacy: Five Ideas That Came From Dartmouth

The Dartmouth Conference did not produce finished results — no papers were published from the conference, no programs were completed, no theorems were proved. But it did produce ideas — conceptual approaches, research programmes, ways of framing problems — that shaped the field for decades.

The Symbolic AI Programme. The most influential idea to emerge from Dartmouth was not a specific result but a general approach: the view that intelligence was fundamentally a matter of symbolic computation, that intelligent behaviour could be achieved by representing knowledge in explicit symbolic structures and manipulating those structures according to explicit rules. This approach — symbolic AI, also called Good Old-Fashioned AI or GOFAI — dominated the field for its first three decades and produced the expert systems of the 1980s, the knowledge representation languages, the logic programming tradition, and much of the foundational theory of AI.

The symbolic approach had its roots in multiple Dartmouth participants: McCarthy’s work on mathematical logic and LISP, Newell and Simon’s work on the Logic Theorist and the General Problem Solver, Minsky’s work on knowledge representation. It was not the only approach discussed at Dartmouth — Rochester’s neural network interests and Solomonoff’s probabilistic approach pointed in different directions — but it was the approach that came to define early AI.

The Heuristic Search Framework. Newell and Simon’s contributions at Dartmouth established heuristic search — the use of domain-specific knowledge to guide the search through a space of possible solutions — as the central technique of AI problem-solving. The Logic Theorist used heuristic search to prove theorems. The General Problem Solver would use it to solve problems across domains. Chess programs, planning systems, and a host of other AI applications would use it in various forms throughout the field’s history. The framework was not new — it had been developing in Newell and Simon’s work before Dartmouth — but the conference gave it exposure and credibility.

The Game-Playing Research Programme. Samuel’s checkers program and Shannon’s analysis of chess established game playing as a central AI research problem. Games were attractive for AI research for several reasons: they were well-defined (the rules were clear), they were challenging (the game trees were enormous), and they provided a clear measure of performance (winning or losing against human opponents). The game-playing research programme produced a continuous series of results over the following decades — from Samuel’s checkers to Deep Blue’s chess to AlphaGo’s Go — each demonstrating what AI could achieve and raising the bar for what would count as genuinely impressive machine intelligence.

The Machine Learning Question. Arthur Samuel’s checkers program — a program that improved with experience, that learned from its own play — introduced the question of machine learning into the AI conversation in a concrete and compelling way. The question of how machines could learn — how they could improve their performance through experience without being explicitly reprogrammed — was identified at Dartmouth as one of the central challenges of AI, and it has remained central to the field ever since. The specific approach that Samuel used, reinforcement learning through rote memory, was one of many that would be developed over the following decades.

The Language Question. The question of how a computer could be programmed to use language — listed first in McCarthy’s original proposal — was identified at Dartmouth as a central challenge. The difficulty of the problem was not fully appreciated in 1956: most of the participants seem to have believed that natural language processing was a challenging but tractable problem that would yield to the right combination of grammar rules and vocabulary. Decades of subsequent research revealed that language was one of the hardest problems in AI — requiring not just grammar and vocabulary but world knowledge, context sensitivity, pragmatic understanding, and a hundred other things that proved extraordinarily difficult to capture in explicit rules. The problem was not solved by the rule-based approaches that dominated early AI. It was eventually addressed, incompletely, by statistical and neural network approaches that the Dartmouth participants were not imagining.

The Institutional Legacy: What the Conference Built

Beyond the ideas, the Dartmouth Conference built something more tangible: a community, and the institutional infrastructure through which that community would sustain and reproduce itself.

The participants did not go back to their universities as isolated individuals. They went back knowing each other — knowing who was working on what, who had complementary approaches, who was worth arguing with, who might be a useful collaborator. The informal network established at Dartmouth became the formal network of the AI research community: the people who were invited to conferences, who reviewed each other’s papers, who trained each other’s students, who built the institutions — the AI labs at MIT and Stanford and Carnegie Mellon — that would dominate the field for the next three decades.

McCarthy and Minsky co-founded MIT’s AI Lab in 1959, three years after Dartmouth. The Lab became the most influential AI research institution in the world, producing an extraordinary concentration of important results and important people over the following decades. Its culture — intensive, competitive, slightly eccentric, deeply committed to the idea that intelligence was a computational process that could be understood and built — was the culture of Dartmouth, amplified and institutionalised.

McCarthy founded the Stanford AI Laboratory in 1963, creating a second pole around which West Coast AI research would coalesce. Stanford’s approach to AI was somewhat different from MIT’s — more concerned with formal logic, with the representation of knowledge in explicit logical structures, with the relationship between AI and mathematical foundations. But the shared identity, the shared vocabulary, the shared research agenda that had been established at Dartmouth made the two labs recognisably part of the same field.

These labs trained the researchers who trained the next generation, who trained the generation after that. The intellectual genealogy of AI — the chains of advisor-student relationships through which ideas, methods, and values are transmitted — runs directly through Dartmouth and through the institutions that the Dartmouth alumni built. The field did not simply begin at Dartmouth in the sense that all the ideas originated there. But the community that would develop those ideas, and all the ideas that followed, was constituted at Dartmouth.

The Overconfidence: Where It Came From and What It Cost

Any honest account of Dartmouth must grapple with the overconfidence — the predictions that proved spectacularly wrong, the timelines that turned out to be off by decades, the gap between what was claimed and what was delivered.

The overconfidence was real, and it mattered. It shaped funding decisions, influenced career choices, set public expectations, and ultimately produced the backlash — the Lighthill Report, the ALPAC Report, the first AI winter — that damaged the field for years. Understanding where the overconfidence came from is important both for understanding the history and for thinking about the similar dynamics that recur whenever AI experiences a new wave of excitement.

Several sources can be identified.

The novelty of the results. When you have just seen a program prove a mathematical theorem — something that had never been done before, something that everyone had assumed required human intelligence — the natural response is to believe that you have crossed a fundamental threshold. If a machine can prove theorems, surely it can do other things that require intelligence. The generalisation feels natural. It turns out to be wrong, because the specific capabilities demonstrated in one domain do not transfer automatically to other domains, and because the difficulty of different cognitive tasks varies enormously in ways that are not obvious in advance.

The lack of a theory of difficulty. The participants at Dartmouth did not have a principled way of assessing how hard different cognitive tasks were, or of predicting how much computational power and how much algorithmic sophistication would be needed to perform them. They could see that proving theorems and playing checkers were harder than arithmetic — they required more sophisticated approaches. But they could not see that natural language understanding and real-world perception were orders of magnitude harder than game playing, or that the gap between laboratory performance and real-world performance would be so consistently and so severely underestimated.

The culture of bold claims. American science in the 1950s, particularly the parts of it connected to defence and national security, had a culture of bold claims and ambitious projections. Researchers who wanted funding needed to convince programme officers and congressional committees that their work was important and that it would produce results. This culture rewarded confident predictions and penalised hedging. The AI researchers who made the most ambitious predictions got the most funding and the most attention. The culture selected for overconfidence.

The isolation of the early community. The AI research community was small in 1956 and would remain relatively small for the next decade. The people who understood what the field was doing were largely the people doing it. There was no mature peer review culture, no established tradition of careful empirical evaluation, no external community of sceptics who could provide the friction that prevents overconfidence from becoming detached from reality. The community was enthusiastic, talented, and largely talking to itself.

The overconfidence cost the field years of progress, millions of dollars in misdirected funding, and the credibility of a generation of researchers who had made promises they could not keep. It was not dishonesty — the participants at Dartmouth believed what they were saying. But belief, however sincere, is not a substitute for evidence. And the evidence, as the 1960s and 1970s unfolded, was not supporting the predictions.

The Shadow of What Was Missing

Dartmouth established the field of AI with considerable success. But the conference also, by what it emphasised and what it neglected, shaped the field’s blindspots — the areas it would underinvest in, the approaches it would overlook, the difficulties it would consistently underestimate.

The most consequential neglect was of learning. The participants at Dartmouth were, with the partial exception of Samuel, oriented toward the explicit representation of knowledge rather than toward the acquisition of knowledge through experience. McCarthy’s LISP and the symbolic AI tradition it supported assumed that knowledge would be put into AI systems by human programmers — that intelligence would be programmed in, not learned. The question of how a machine could acquire the vast, complex, context-sensitive knowledge it needed to navigate the real world — a question that would prove to be one of the hardest in all of AI — was largely set aside.

This neglect had consequences that lasted decades. The symbolic AI systems of the 1960s and 1970s were brittle precisely because they were hand-coded — they contained only the knowledge that their programmers had explicitly provided, and that knowledge was always inadequate to the full complexity of the real world. The systems broke down at the boundary of their explicitly programmed knowledge, because they had no way to extend that knowledge through experience.

The learning-based approaches that would eventually prove most productive — neural networks, statistical machine learning, reinforcement learning — were not absent from the Dartmouth conversation, but they were not the dominant paradigm. Rochester’s neural network interests and Solomonoff’s probabilistic approach pointed in the right direction. But the field that emerged from Dartmouth was oriented toward the symbolic, toward the explicit, toward the programmed — and that orientation persisted, with diminishing returns, for thirty years.

The second significant neglect was of perception. The problem of getting a machine to perceive the real world — to see, to hear, to understand the sensory environment in which intelligent action had to take place — was not central to the Dartmouth vision. The problems that the conference focused on — theorem proving, game playing, language use, planning — were problems that could be tackled with symbols and rules, without any need to deal with the messy complexity of real-world perceptual data.

This neglect was also consequential. Perception turned out to be one of the hardest AI problems — so hard that it was still largely unsolved thirty years after Dartmouth. The breakthrough on perceptual AI — the development of deep convolutional neural networks that could match human performance on image recognition — came in 2012, more than half a century after Dartmouth, and came from an approach — deep learning, large datasets, GPU computation — that was not being discussed at the 1956 conference.

The third neglect was of common sense. The real world is not a theorem-proving domain or a chess game. It is an environment of extraordinary complexity and ambiguity, navigated using a vast body of implicit, tacit, common-sense knowledge that humans acquire through childhood development and daily experience. Getting a machine to represent and use this knowledge — to know that water flows downward, that people usually eat with utensils, that a dropped object will fall, that a person who is crying is probably sad — turned out to be extraordinarily difficult, harder in some ways than any of the problems that Dartmouth foregrounded.

The common-sense knowledge problem was identified, as a problem, by AI researchers working in the tradition Dartmouth established. Douglas Lenat’s Cyc project — a multi-decade effort to manually encode common-sense knowledge into a machine-readable form — was a direct response to this challenge. But the manual encoding approach was doomed to fail, because the quantity of common-sense knowledge was too vast, its structure too complex, its contextual sensitivity too fine-grained for explicit encoding to capture. The eventual approach that made progress on common-sense reasoning — training large language models on vast quantities of human text — was not imagined at Dartmouth.

What Dartmouth Got Right

It is easy, with the benefit of seventy years of hindsight, to focus on what Dartmouth got wrong — the overconfidence, the neglected approaches, the blindspots that cost the field decades. But Dartmouth also got important things right, and those deserve acknowledgment.

The conjecture was right. The founding claim — that intelligence can in principle be described precisely enough to be simulated by a machine — was a bet, and it was a good bet. The evidence accumulated over the following decades, and particularly over the past decade, strongly suggests that at least a great deal of what human intelligence does can be replicated, or approximated, or closely approached, by machine systems. The bet has not been definitively won — the hard questions about consciousness and genuine understanding remain open — but it has not been definitively lost, and the progress has been real and substantial.

The research programme was right. The problems that Dartmouth identified — language, learning, perception, reasoning, planning — were the right problems. They are still the central problems of AI. The approaches to those problems were inadequate, but the problems were correctly identified. A field that works on the right problems, even with inadequate methods, will eventually develop better methods. A field that works on the wrong problems will not.

The community was right. The decision to create a named field — to give AI an identity, a community, a set of institutions — was correct. Science happens in communities, and the community that Dartmouth created was, despite its flaws, extraordinarily productive. The researchers it trained, the institutions it built, the culture of ambitious, rigorous, self-critical inquiry it established — these were genuine contributions to human knowledge.

The Summer in the Long View

Standing at the distance of seven decades, the Dartmouth Conference looks both more and less momentous than it has sometimes been presented.

Less momentous, because the ideas that were foundational to the field — Turing’s computability theory, Shannon’s information theory, the McCulloch-Pitts neural model, Wiener’s cybernetics, von Neumann’s computer architecture — were mostly in place before Dartmouth. The conference did not create the intellectual foundations. It named them, organised them, gave them a shared identity and a shared programme.

More momentous, because names and communities matter more than is usually acknowledged. The history of science is full of fields that had the right ideas but failed to cohere into productive communities — that remained scattered, disorganised, underfunded, unable to recruit students or attract institutional support. AI could have remained a scattered collection of individual projects, each working on a piece of the problem without the network effects that come from being part of a recognised field. The Dartmouth Conference prevented that. It created the field, in the sociological and institutional sense, and everything the field has subsequently achieved depends on that creation.

The summer of 1956 was not the summer intelligence was invented, or even the summer the project of building machine intelligence was first seriously pursued. It was the summer that project got a name, a community, and a mission. And that was enough. That was what was needed. The rest followed.

Further Reading

  • The original Dartmouth proposal (1955) — McCarthy’s proposal to the Rockefeller Foundation is available online. Short, readable, historically essential. Read it in the original.
  • “Machines Who Think” by Pamela McCorduck — The most comprehensive popular history of AI, with a vivid account of Dartmouth and its participants.
  • “The Innovators” by Walter Isaacson — Provides essential context for Dartmouth through the broader story of the people who invented the digital age.
  • “Hackers: Heroes of the Computer Revolution” by Steven Levy — Captures the culture and spirit of the early AI community at MIT, the direct institutional descendant of Dartmouth.
  • “Artificial Intelligence: A Modern Approach” by Stuart Russell and Peter Norvig — The standard AI textbook, whose opening historical chapter provides a rigorous and balanced account of the field’s founding.

Next in the Articles series: A7 — The First AI Programs: Teaching Machines to Play Games — The exhilarating early years when Arthur Samuel’s checkers program learned to beat its creator, when chess programs first challenged amateur players, and when every new demonstration felt like proof that general machine intelligence was just around the corner. What those early programs actually did, why they were so impressive, and why the gap between their performance and genuine intelligence was wider than it appeared.

Minds & Machines: The Story of AI is published weekly. If the founding story of AI — with all its ambition, its insight, and its spectacular overconfidence — resonates with conversations you are having today about the field, share it with someone who would find the parallel illuminating.