The Birth of AI: Dartmouth, Symbolic Systems, and Early Optimism

Part of The Long Human Road to AI.

Before the field had a name, the question was already in the air. In 1950 Alan Turing asked whether a machine could imitate human reasoning well enough to fool an interrogator. Digital computers were just becoming reliable, and a scattered community of mathematicians, engineers, logicians, and psychologists was already trying to make them act intelligent. The “birth” of artificial intelligence, then, was not the first moment anyone imagined a thinking machine. It was the moment a small group of researchers drew a map around scattered settlements and gave the whole territory a name.

A Field Gets a Name

In the summer of 1956, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon hosted a two-month workshop at Dartmouth College. Their 1955 proposal argued that learning and other features of intelligence could, in principle, be described precisely enough for a machine to simulate them. The workshop gave that ambition its durable label: “artificial intelligence.”

Point C1 Dartmouth named and consolidated AI as a research field, but it did not originate all machine-intelligence work.

The proposal was bold, but it was also a funding document and a recruiting signal. It gathered people who had been working on games, proofs, languages, and neural networks under one tent. In that sense Dartmouth was less a single invention than a field-forming moment: it turned separate projects into a shared conversation with shared institutions.

This does not mean the conversation started there. Turing’s 1950 paper, cybernetics, information theory, and wartime work on computation all fed into the gathering. The name helped the field become legible to funders, universities, and the public; it did not erase what came before it.

The Conjecture

The central bet of early AI was that intelligence could be made operational. If a task could be represented as symbols, rules, and goals, then a computer could search through possible moves, proofs, or actions and choose one that looked best.

Point C2 Early AI treated reasoning as symbolic manipulation and search.

This was a powerful idea. It translated something as abstract as “thinking” into procedures a machine could run: represent the problem, define the legal moves, pick heuristics to cut the search space, and execute. The wager was not that the machine understood the world the way a person does, but that enough intelligence-like behavior could be built from formal structures and careful search.

Proofs, Plans, and Programs

The researchers set out to make the conjecture concrete. Allen Newell and Herbert Simon’s Logic Theory Machine, presented in 1956, searched for proofs in Principia Mathematica. The General Problem Solver, developed in the late 1950s and early 1960s, tried to apply means-ends analysis to a wider range of puzzles. John McCarthy’s 1959 “Programs with Common Sense” proposed a system, the Advice Taker, that would use a formal language to represent everyday knowledge and draw conclusions from it.

Point C3 Early demonstrations were impressive but bounded: they worked inside formal or carefully prepared worlds.

These systems were genuinely startling. A machine that could prove theorems or plan moves looked like the opening act of a much larger drama. Yet each success was also a constrained one. Logic Theorist operated on already-formalized mathematics. GPS needed its problems encoded in a form it could search. The Advice Taker remained a proposal, not a running program. The demonstrations showed that symbol manipulation was possible; they did not show that open-ended human intelligence had been captured.

Why Optimism Made Sense

From the vantage point of the late 1950s, the optimism was not absurd. Electronic computers were new and fast. Formal logic seemed to capture the structure of valid thought. Wartime and postwar institutions were pouring money into computation. Small demos could feel like harbingers of a much larger future.

Point C4 Early AI optimism was part technical, part institutional, and part public narrative.

The 1955 Dartmouth proposal promised dramatic advances. Press accounts amplified the excitement. Frank Rosenblatt’s perceptron work, covered in outlets like the Cornell Chronicle, added a parallel thread of neural optimism. Funding, prestige, and popular imagination began to reinforce one another. What looked plausible from inside the period was, in retrospect, a loop in which real progress, institutional ambition, and storytelling each made the others larger.

Why Optimism Was Not Enough

Even as symbolic AI claimed center stage, the field was never a single lineage. Turing’s 1950 question about machine intelligence sat beside cybernetics, game-playing programs, information theory, and Rosenblatt’s perceptron. The Dartmouth-centered story is useful, but it becomes misleading if it is treated as the whole story.

Point C5 The early field included multiple lineages, including symbolic reasoning, cybernetics, neural approaches, game-playing, and machine-intelligence philosophy.

The limits that would later humble the field were also visible early, if anyone looked closely. Real-world perception, common sense, open-ended language, and messy environments resisted clean formalization. Symbolic systems excelled when the rules were explicit and the world was small. They struggled when the task required the kind of flexible, contextual judgment humans take for granted.

The Durable Legacy

Early symbolic AI did not deliver general machine intelligence. What it delivered was a set of enduring ideas: search, representation, planning, formal languages, and the habit of turning intelligence into testable systems. Lisp, born in 1958, would outlast many of the ambitions that surrounded it. The critique of over-simple origin myths, advanced by historians such as Stephanie Dick, reminds us that the field’s identity was contested from the start.

The real lesson of the birth of AI is not that researchers solved intelligence in the 1950s. It is that they learned how much of intelligence would resist being turned into symbols, rules, search, and programs. That learning set the stage for the next chapter: the winters, expert systems, and cycles of promise and disappointment that followed.