Symbolic AI
Symbolic AI is the paradigm of building intelligent behavior from explicit, human-readable symbols (words, logical predicates, frames, graph nodes) and operations that manipulate them according to rules, so that reasoning is carried out by deduction, search and planning over a represented model of the world rather than by weights adjusted from data. It dominated artificial intelligence from the 1956 Dartmouth workshop, where John McCarthy's term "artificial intelligence" was adopted, until statistical learning displaced it, and the philosopher John Haugeland nicknamed it "good old-fashioned AI," or GOFAI, in 1985.
Its founding claim is that thinking is symbol processing. Allen Newell and Herbert Simon's Logic Theorist (1956) proved theorems from Whitehead and Russell's Principia Mathematica by searching through logical derivations, and their General Problem Solver (1957 onward) separated a generic search procedure from task-specific knowledge. In 1976 they stated the physical symbol system hypothesis: a physical system that manipulates symbol structures has the necessary and sufficient means for general intelligent action. Under this view a mind and a computer are the same kind of thing, which justified decades of work on representation languages such as LISP (McCarthy, 1958), Prolog (Alain Colmerauer and colleagues, 1972), semantic networks, frames and description logics.
The paradigm's strength is that its knowledge and reasoning are inspectable and exact. A planner such as STRIPS (Fikes and Nilsson, 1971) produces a sequence of actions whose every step can be traced to a stated precondition and effect. A theorem prover either finds a proof or does not, and a proof can be checked independently. Because rules are written by people, systems can encode prior knowledge, honor hard constraints and explain their conclusions without needing large datasets. These properties are why symbolic methods never disappeared: SAT and SMT solvers verify chips and software, constraint solvers schedule airlines and factories, and proof assistants such as Lean and Coq certify mathematics.
Its weaknesses explain why learning took over. Hand-written knowledge is expensive and never complete; Cyc, the ambitious project Douglas Lenat began in 1984 to encode common sense, illustrates both the scale of the effort and its open-endedness. Early systems suffered from the frame problem (named by McCarthy and Hayes in 1969), the difficulty of stating what stays unchanged when an action occurs, and from the qualification problem, since almost every rule has exceptions. Perception was the harshest test: no one could write rules that recognized a cat in pixels, and the symbol grounding problem (Stevan Harnad, 1990) asked how symbols acquire meaning at all. Shortfalls like these, together with the funding collapses described under AI winter, pushed the field toward statistical methods.
Symbolic AI is distinct from its two nearest neighbors in this glossary. An expert system is one deployed product family built within the paradigm, packaging domain knowledge into a rule base and an inference engine, while a rule-based system is the narrower if-then mechanism that many such products use. Symbolic AI is the whole stance: representation, search, logic, planning and knowledge graphs, whether or not any rules in the narrow sense are involved. Its techniques also run well beyond rules, as in graph search algorithms like A*, ontologies and knowledge graphs that power search engines, and probabilistic programs.
Today the live question is how to combine symbols with learning. Neurosymbolic approaches use neural networks to handle perception and fuzzy language while delegating exact steps to symbolic machinery. DeepMind's AlphaGeometry (Trinh and colleagues, 2024) paired a language model that suggests constructions with a symbolic deduction engine that checks proofs, solving a large share of olympiad geometry problems. Tool-using language models likewise call calculators, databases, code interpreters and formal verifiers, effectively outsourcing precision to symbolic components. A frequent misconception is that symbolic AI was refuted by deep learning; it was instead outcompeted on perception and open-ended language, while remaining the stronger tool wherever guarantees, auditability or explicit constraints matter.
Read symbolic AI as the bet that intelligence can be written down: its enduring achievement is exactness and explanation, its enduring limit is that the world is too large and messy to be written down by hand, and the most promising modern systems pair learned perception with symbolic checking rather than choosing one camp.