Symbolic AI
Symbolic AI writes knowledge down as readable symbols and if-then rules, then reaches answers by applying logic and search to them.
Symbolic AI stores knowledge as symbols a person can read, such as parent(ada, ben), and works on them with rules. Allen Newell and Herbert Simon called such a program a physical symbol system. It holds symbols, which are physical patterns, combined into expressions, plus processes that turn expressions into new ones. In their Turing Award lecture they argued that a physical symbol system is all a machine needs for general intelligence, and that anything generally intelligent must be one. Such a system, they said, solves a problem by searching: it builds candidate symbol structures and keeps altering them until one counts as an answer. Rule-based systems of this kind are often called GOFAI, short for good old-fashioned AI.
The idea is as old as the field’s name. The 1955 proposal for the Dartmouth summer project started from a bold guess: that learning and intelligence could, in principle, be pinned down so exactly that a machine could copy them. It also guessed that a lot of human thinking is shuffling words by the rules of reasoning. The same proposal also listed neuron nets as a topic. At the 1956 conference, Newell and Simon showed the Logic Theorist, which proved elementary theorems in propositional logic.
The rival camp, connectionism, says intelligence does not come from shuffling symbols at all, but from brain-like processing in artificial neural networks. A neural network does not store written rules. Training sets its weights, the values it multiplies its inputs by. McCarthy saw learning from experience as the strong suit of neural-net approaches. In 2007 McCarthy named common sense as the area where AI was farthest from human level. He called the Cyc system’s collection of common sense facts “large but spotty”.
The two camps are no longer strict rivals. Serious systems now often combine logic with neural networks. AlphaGeometry, described in Nature in January 2024, pairs a neural language model with a symbolic deduction engine. When the engine gets stuck, the language model adds a new point, and the engine tries again. It solved 25 of 30 olympiad geometry problems within the time limit, against 10 for the previous best method. Its proofs are human-readable.
A program that reasons needs a way to store facts, draw new ones from them and choose among many options.
Follow one question through a small, illustrative family-tree program.
- 1 · writePeople write facts and rules as symbol expressions, such as parent(ada, ben) and a rule defining grandparent.
- 2 · askA question sets the test an answer must pass, here whether grandparent(ada, cal) holds.
- 3 · matchThe inference engine finds a rule whose if parts match known facts, filling the rule's variables with names.
- 4 · deriveThe rule's then part becomes a new fact, which answers the question.
- 5 · explainThe system can show the facts and rule it used, so a person can check each step.
Nothing here was learned from data: people wrote every fact and rule, so the answers are only as good as that writing.
| Who | What they ask | What it works with |
|---|---|---|
| Maths research team | “Can a computer find an olympiad geometry proof we can check line by line?” | Geometry rules applied by a symbolic deduction engine |
| Biomedical data team | “Does our ontology classify every term consistently?” | An ontology written in a description logic |
| Card payments team | “Should we accept this card purchase?” | Facts about the card owner, the item and the shop, sorted into fixed categories |
| Game programmers | “Which chess move looks best from here?” | Possible moves, examined by search |
- Conclusions come from clear rules, so the reasoning can be explained step by step.
- Answers can be checked by machine. Every AlphaGeometry solution was verified by computer.
- Search lets it explore many options, such as chess moves or proof steps, in an orderly way.
- Deduction suits tasks where an answer must follow from stated premises, such as proving theorems.
- Someone has to write the knowledge down, and what experts know gets squeezed into whatever format the program was built around.
- Written-down common sense can be patchy. In 2007 McCarthy called the Cyc system's collection of common sense facts “large but spotty”.
- Ordinary logic cannot withdraw a conclusion. Told that birds fly, it needs extra machinery to accept that a penguin does not.
- On big, tangled problems, a deduction engine working alone can crawl and struggle to adapt.
Sources used
This explainer is written in original language. The links below support its factual claims.
- paperComputer Science as Empirical Inquiry: Symbols and Search (1975 Turing Award lecture), Communications of the ACM 19(3), 1976, Newell and Simon, Association for Computing Machinery · read 27 Sept 2026
- paperA Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, McCarthy, Minsky, Rochester and Shannon (Stanford copy) · read 27 Sept 2026
- paperWhat is Artificial Intelligence?, John McCarthy, Stanford University · read 27 Sept 2026
- articleArtificial Intelligence, Stanford Encyclopedia of Philosophy · read 27 Sept 2026
- articleThe Chinese Room Argument, Stanford Encyclopedia of Philosophy · read 27 Sept 2026
- officialAlphaGeometry: An Olympiad-level AI system for geometry, Google DeepMind · read 27 Sept 2026
- paperSolving olympiad geometry without human demonstrations, Trinh et al., Nature · read 27 Sept 2026
- docsMachine Learning Glossary, Google for Developers · read 27 Sept 2026