The History of Symbolic AI

By Victor Hoang · Updated Oct 22, 2026

In the winter of 1955 to 1956, Newell and Simon have an idea. A machine could prove logic theorems like a person: trying moves, setting subgoals, discarding dead ends. No machine runs it yet. They simulate it by hand, with family members and students as the parts.

Symbolic AI builds intelligence by letting a computer manipulate symbols according to rules. As a research program it began in 1955 and 1956 with Allen Newell, Herbert Simon and J. C. Shaw's Logic Theorist. It proved theorems from Whitehead and Russell's Principia Mathematica. The idea is far older: philosophers spent more than two thousand years trying to make reasoning rote. The throughline is a rise, a fall and a return. Symbolic methods dominated the field's first three decades. They lost ground when they couldn't scale or absorb common sense. Now they're wired back into systems built around neural networks.

Out of that improvisation came the first generation of "thinking" programs, and an argument about what thinking is that still runs through the field. For the wider story, see The History of AI and the AI timeline.

Timeline at a Glance

The Paper Years (4th century BCE-1913)

Logic turns from argument into notation

Imagine a medieval scholar turning a paper disc inside a paper disc, so that letters on one ring line up with letters on another, and every alignment is a statement to be examined. That is the picture behind Ramon Llull's Ars. According to the Stanford Encyclopedia of Philosophy, Llull wrote the first version in 1274 and kept reworking it until the Ars generalis ultima of 1305 to 1308. Its figures used lettered terms, with 54 terms in the later "ternary" phase, and its Fourth Figure was three concentric circles rotated to produce combinations. Llull meant it as a method for finding arguments and reaching necessary conclusions, aimed at persuading people of other faiths from shared principles. It was a combinatorial machine for reasoning, built of paper.

Llull is a late stop on a long road. The road starts with Aristotle's Prior Analytics, which the Stanford Encyclopedia calls the earliest formal study of logic that we have. Aristotle's syllogism takes two premises that share one term and derives a conclusion, and he checks every pairing of premises, listing 14 valid moods in three figures. The key move is that validity depends on the pattern, written with letters in place of subjects. If you can reason by pattern, you can in principle reason by machine.

The seventeenth-century philosopher Gottfried Leibniz pushed that idea furthest. The Stanford Encyclopedia's account of his influence describes his plan for a characteristica universalis, a universal sign language in which concepts are broken into elements so that they can be handled by calculation, and a calculus ratiocinator, a calculus for reasoning. Errors in reasoning would show up as wrong combinations of signs, and a dispute could be settled the way a sum is settled. The same source records a nineteenth-century critic's complaint that the analysis of concepts it needed was not yet possible.

Two nineteenth-century books made the dream concrete. George Boole's 1854 Laws of Thought turned logic into an algebra, and the Stanford Encyclopedia notes that later logicians judged it built in a Leibnizian spirit even though Boole had not read Leibniz's anticipations when he wrote it. Gottlob Frege's Begriffsschrift of 1879, subtitled a formula language of pure thought modelled on arithmetic, added an analysis of generality, the "every" and "some" of ordinary argument. Frege described his own work as a step toward Leibniz's ambitions. Then Alfred North Whitehead and Bertrand Russell published Principia Mathematica in 1910, 1912 and 1913, defending the thesis that mathematics can be reduced to logic by deriving it in fully formal steps.

Be clear about what this is. None of these people built an AI, and nobody before the 1950s called it that. They are precursors: they proved that reasoning could be written in a formal notation whose steps do not depend on what the symbols mean. That is the premise symbolic AI inherited whole. What changed in the twentieth century was that a machine could execute the notation.

For teams today: the oldest lesson is that anything you can state as explicit steps can be checked, audited and repeated. That is still why a written approval rule is easier to trust than a verbal one.

The Logic Theorist Years (1955-1959)

Four ideas arrive at once: search, symbols, a language, a name

At the end of August 1955, John McCarthy of Dartmouth, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories put their names to a proposal. The document is dated 31 August 1955 and asks for a two-month, ten-man study of artificial intelligence during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. Its working conjecture is that every aspect of learning or any other feature of intelligence can in principle be described precisely enough for a machine to simulate it. The authors also suggest that much of human thought consists of manipulating words according to rules, and propose an artificial language a computer could be programmed to use. The 1956 Dartmouth workshop that followed gave the field its name.

Newell and Simon were the ones who arrived with something running, or close to it. Jeff Shrager's study of the Logic Theorist's source code records Edward Feigenbaum's recollection of the moment in January 1956, when Simon told his class that over Christmas he and Newell had "invented a thinking machine". The program was specified in a 1956 RAND report and simulated by hand, with relatives and students acting as program components. Only with the implementation of the IPL-II language on RAND's JOHNNIAC computer in late 1956 and early 1957 did machine-executable proofs appear, and the result usually quoted, 38 of the first 52 theorems in chapter 2 of Principia, comes from those 1957 runs. For Theorem 2.85 it found a shorter proof than the published one, and Simon showed it to Russell, who responded with delight.

How did it work? Nils Nilsson's history The Quest for Artificial Intelligence explains that the Logic Theorist transformed symbol structures built from Whitehead and Russell's axioms until one matched the theorem to prove. Because the possible transformations multiply quickly, the program searched, guided by what Newell and Simon called heuristics. They defined a heuristic as a process that may solve a problem but offers no guarantee of doing so. Nilsson also records that Marvin Minsky, for his part, found heuristic search obvious enough that the Logic Theorist did not impress him.

At the end of the decade the symbolic approach got a method, a language and a manifesto. In 1959, at a conference in Paris, Newell, Shaw and Simon presented the General Problem Solver. Nilsson describes its core mechanism, which they called means-ends analysis: compute the difference between the current state and the goal, find an operator relevant to that difference, and treat any unmet precondition of the operator as a new subproblem. Newell and Simon also claimed the program was a theory of human problem solving, and they wrote that it blurred the line between doing a task and simulating how people do it.

McCarthy, meanwhile, was working on the language. In his April 1960 Communications of the ACM paper, he explains that Lisp was developed for the IBM 704 by the MIT artificial intelligence group to support experiments with a proposed system called the Advice Taker, first proposed in November 1958. The main requirement was a system for manipulating expressions representing formalized declarative and imperative sentences, so that the Advice Taker could make deductions. In the Advice Taker paper, McCarthy contrasts it with the Logic Theorist and Gelernter's Geometry Program, where the heuristics were all embedded in the program: he wanted a system whose "behavior will be improvable merely by making statements to it". The idea of telling a machine facts instead of re-coding it is the ancestor of every knowledge base. For the language's own story, see The History of Lisp in AI and the broader history of programming languages.

This era solved a real puzzle: it showed that a machine could do something that looked like reasoning. It created a new problem too. Search over symbol structures grows explosively, and the knowledge to prune it had to come from somewhere.

For teams today: a demo that proves a hard problem can be solved once is not a system that solves it routinely. The Logic Theorist ran on hand-picked theorems, and the gap between that and general use took decades to measure.

The Confident Years (1960-1972)

Symbols and search are expected to scale to everything

By the early 1960s, symbolic AI was the main road. Nilsson's account of this period describes programs that played games, proved geometry theorems and manipulated algebra, all built on the same recipe: represent the problem as symbol structures, apply transformations, and guide the search with heuristics. Researchers made bold forecasts. Lighthill, reviewing the period later, wrote that in no part of the field had the discoveries made so far produced the major impact that was then promised, a line Nilsson quotes.

The trouble was already visible. Nilsson notes that early researchers knew about combinatorial explosions, which is why they cared about heuristics. But heuristics, he writes, do not change the exponential character of search; at best they reduce the branching factor. His example: cutting the branching factor from 10 to 4 and searching to depth 20 still means roughly a trillion nodes. A method that handled puzzles and logic exercises could stall on real-world problems, where the number of possibilities is far larger.

A second problem was quieter. Programs like the General Problem Solver knew the form of a task but not the facts of a domain. Real tasks needed facts about the world, and the question of how to get that knowledge into a program, and keep it consistent, would shape the next twenty years. The expert systems of the 1970s and 1980s were the response.

For teams today: a method that works on a toy problem can fail on a real one for reasons that are arithmetic, not opinion. Ask for the scale test before the pilot.

The First Winter (1972-1980)

Lighthill's survey cools the money, and the theory gets its name

In July 1972, the Cambridge mathematician Sir James Lighthill dated a survey of AI written for Britain's Science Research Council. The council had been receiving more and more applications in the field, and its chairman commissioned an independent report from someone outside it. According to the Edinburgh AIAI archive, the report is titled Artificial Intelligence: A General Survey, and the archive also holds a BBC-televised debate of June 1973, the "Lighthill Controversy", at the Royal Institution with Lighthill, Donald Michie, Richard Gregory and John McCarthy. Nilsson summarizes the survey's verdict: it concluded that AI search techniques, which worked on toy problems, would not scale to real problems because of the combinatorial explosion. It led to a substantial curtailment of AI research in the United Kingdom, including the work of Donald Michie's group in Edinburgh. The 1970s were a hard decade, but Nilsson adds that applications work kept funding alive and that excitement about expert systems then built toward the mid-1980s.

Critics from outside the field were also at work. Nilsson lists the philosopher Hubert Dreyfus, whose What Computers Can't Do appeared in 1972, among the outsiders who challenged the premise itself. This is where the symbolic approach got its sharpest statement. On 20 October 1975, at the ACM annual conference in Minneapolis, Newell and Simon gave their Turing Award lecture, and the paper appeared in CACM's March 1976 issue. It stated the Physical Symbol System Hypothesis: "A physical symbol system has the necessary and sufficient means for general intelligent action." A physical symbol system, they wrote, is a set of symbols, which are physical patterns, plus processes that produce an evolving collection of symbol structures. They paired it with a second claim, that such a system exercises intelligence in problem solving by search, and said the ideas were already present, though unnamed, in the Logic Theorist in 1955.

Read in the context of the 1970s, the lecture was a closing argument. It took a working method and elevated it into a claim about the nature of intelligence, in the same years the method's limits were becoming plain. Nilsson would later write that this claim was the primary rationale for the approach Haugeland named GOFAI.

For teams today: a winter in funding can coexist with a good idea. Lighthill's critique was about scale and cost, and the same questions are fair to ask of any AI pilot.

The Naming and the Second Winter (1980s-1990s)

Rules and common sense get a big bet, and a name that sticks

In the 1980s the symbolic method found paying customers. Rule-based expert systems packaged a specialist's judgment into software and spread through business; Nilsson calls roughly 1975 to 1985 "boomtimes" for AI. That story, from DENDRAL and MYCIN to XCON and the downturn, is told in The History of Expert Systems, and this article does not repeat it. What matters here is how it touched the symbolic idea. Expert systems proved that explicit rules could do paid work, and also showed what that cost: someone had to write and maintain the rules.

The Advice Taker's old question came back as a bet in 1984. According to Nilsson, Stanford professor Douglas Lenat concluded that common-sense knowledge was needed for many applications, especially natural language understanding. In the fall of 1984 he left academia for the newly formed Microelectronics and Computer Consortium (MCC) in Austin, Texas, to assemble a team that would hand-enter "a couple of million important facts and rules of thumb" into a program called Cyc, short for encyclopedia. Lenat originally estimated a couple of million general assertions; Nilsson later reports he believed the number was closer to 200 million. The team entered knowledge by hand for about ten years at MCC, and in 1994 Lenat founded Cycorp to commercialize the work. Cyc is the clearest test of the symbolic thesis at scale: can common sense be written down as rules? A modern descendant of that idea is the knowledge graph, which stores facts as linked entities.

The approach got its name at the moment it was being challenged. John Haugeland's 1985 book, Artificial Intelligence: The Very Idea (MIT Press), called it "good old-fashioned artificial intelligence", abbreviated GOFAI, as Nilsson notes. Nilsson says GOFAI seemed to be running out of steam during the 1980s, which made it vulnerable to challenges from AI researchers themselves. Some, he writes, pursued "connectionist" or "brain-style" computation, the line of work behind today's neural networks, to contrast it with GOFAI. In the same chapter Nilsson recalls that Drew McDermott, a Yale professor who had been among the logicists, began to doubt the role of logic in AI. The doubts came from inside the camp.

Then the second winter arrived, covered in the expert-systems article: funding and company closures followed the end of the 1980s boom, and researchers turned to new ideas, including learning from data (see machine learning).

For teams today: a knowledge base is a product, not a project. Cyc's own founder revised his estimate of the knowledge needed from millions of assertions to hundreds of millions. Budget for the size you will discover, not the size you planned.

Eclipse and Comeback (1990s-present)

Learned models take over, and symbols return as a check

After the second winter, much of the field’s energy moved from writing rules to learning from data. Symbolic methods did not disappear; they went inside other software, as with business rules, and into specialized fields where proofs and constraints matter. But the research energy moved toward machine learning and, later, large language models, which are fluent but whose reasoning cannot be inspected rule by rule.

The weaknesses of each camp are now complementary. Neural models learn from data but can be confidently wrong, and symbolic systems can be checked but were costly to build by hand. IBM Research describes neuro-symbolic AI as combining statistical methods like machine learning with human-like symbolic knowledge and reasoning, and lists work from the MIT-IBM Watson AI Lab on making commonsense AI decisions explainable. That is institutional framing from a company with an interest in the approach, so read it as a statement of aims, not a verdict.

The clearest public demonstration so far is AlphaGeometry. On 17 January 2024, Google DeepMind's announcement, published alongside a Nature paper, described a system that pairs a neural language model with a symbolic deduction engine. The symbolic engine deduces new statements about a diagram, and if it gets stuck the language model proposes a construct, such as a point or line, that opens new deductions, and the loop repeats. On a 30-problem benchmark of olympiad geometry it solved 25, against 10 for the previous best method and 25.9 for the average human gold medalist, DeepMind reports. The two halves map onto this history. The language model supplies the guess that Newell and Simon called a heuristic, and the symbolic engine supplies the proof-style check that Aristotle, Frege and Russell worked toward. The newest reasoning models take a different route to similar goals, but they raise the same question symbolic AI always asked: how do you know the steps are right?

The history is not finished. Whether hybrids become the standard design is an open question, and the evidence so far is a set of strong results in structured domains such as geometry. The one safe claim is that the old tools are back in use alongside the new ones, in the artificial intelligence systems of the present.

For teams today: use learned models for the messy middle and explicit rules or checkers for what must be right. If a decision needs to be explained to an auditor, a rule or a verifiable step is cheaper than an argument.

Key People

Key Facts

• The Dartmouth proposal is dated 31 August 1955 and was written by McCarthy, Minsky, Rochester and Shannon.
• The Logic Theorist proved 38 of the first 52 theorems in chapter 2 of Principia Mathematica, in runs on JOHNNIAC in 1957.
• McCarthy's Lisp paper appeared in CACM in April 1960 and says Lisp was built to support the Advice Taker.
• Newell and Simon delivered their Turing Award lecture on 20 October 1975, published in CACM in March 1976.
• Lighthill's survey is dated July 1972, with a BBC debate recorded in June 1973.
• AlphaGeometry solved 25 of 30 olympiad geometry problems, against 10 for the previous best method.

The Pattern Behind the History

Across six eras, the same four moves repeat. We call the sequence the Formalize-Search-Stall-Hybridize Loop. It is our synthesis of the history above, not a research finding.

1. Formalize. Aristotle, Boole, Frege and Russell each wrote reasoning in a notation that ignores what the symbols mean.
2. Search. The Logic Theorist and the General Problem Solver applied the notation by trying moves and pruning with heuristics.
3. Stall. Combinatorial explosion in 1972, and the cost of entering common sense by hand in the 1980s, capped each push.
4. Hybridize. Each time, the symbolic idea survived by pairing with a newer tool: rules inside business software, and now a deduction engine alongside a language model.

What It Teaches Teams Today

• Write down the rule before you automate it. Explicit steps can be checked, and symbolic AI began as that discipline.
• Test at scale early. Heuristics reduce the search; they do not remove the explosion.
• Treat knowledge as a living asset. Cyc's estimate grew a hundredfold, and someone has to own the update.
• Match the tool to the failure cost. Where a wrong answer is expensive, add a rule or a checker around the model.
• Ask what the old critics asked. Does it scale, who maintains it, and how would you know it is right?

Symbolic AI taught the field that reasoning can be written down, searched and checked, and that writing it all down never ends. As learned models spread, the question it asked in 1955 is still the useful one: can we see how the answer was reached?