What was symbolic AI?

Artificial intelligence began as an academic field in the summer of 1956, at a small conference at Dartmouth College in New Hampshire. The Stanford Encyclopedia of Philosophy notes that the term “artificial intelligence” was coined there, and that the attendees included John McCarthy, Marvin Minsky, Claude Shannon, Allen Newell, and Herbert Simon.^[1] At that same meeting, Newell and Simon showed a program called Logic Theorist, which could prove elementary theorems in logic. It was treated as a remarkable result, “agreed by the attendees … to be a remarkable achievement.”^[1]

The optimism that followed ran ahead of the results. Alan Turing had predicted his own imitation-game test would be passed by the year 2000.^[1] It was not: as the Stanford Encyclopedia notes decades later, “the most articulate of computers still can’t meaningfully debate a sharp toddler.”^[1] That gap between an early, confident timeline and what the field could actually deliver would repeat itself more than once in the decades that followed.

The approach those founders took is now called symbolic AI. The idea is that thinking is the manipulation of symbols. You give the machine facts written as symbols, plus rules for combining them, and reasoning is what happens when the machine applies the rules to the facts. This grew out of classical deductive logic, where a conclusion follows with certainty from a set of premises.^[1]

Newell and Simon later put this belief into a single sentence. Their physical symbol system hypothesis, from a 1976 paper, states that “a physical symbol system has the necessary and sufficient means for general intelligent action.”^[2] In plain terms: a system that stores symbols, combines them into structures, and manipulates them with processes has everything it needs to be intelligent, and nothing else is required.^[2] That was the bet the field made for its first three decades.