Definition
Symbolic AI — sometimes called Good Old-Fashioned Artificial Intelligence (GOFAI) — is the paradigm that dominated the field from its founding in the 1950s through the late 1980s. Its central claim is that intelligent behavior can be produced by manipulating explicit symbols according to formal rules. The mind, on this view, is essentially a sophisticated symbol manipulator, and a sufficiently well-programmed computer can replicate intelligent behavior by implementing the same kind of rule-governed symbol processing.
The paradigm rests on two commitments. First, that knowledge about the world can be encoded in explicit, discrete representations: propositions, semantic networks, frames, production rules, ontologies. Second, that reasoning is a formal process that operates over those representations — deduction, search, pattern matching — rather than a statistical or connectionist process that operates over numerical weights. Both commitments are philosophically substantive, and both have been challenged.
The early achievements of symbolic AI were real and striking. Programs could prove mathematical theorems, play chess at high levels, diagnose diseases from symptom lists, parse natural language sentences, and schedule airline routes. Expert systems, which encoded the knowledge of human specialists in large rule bases, were commercially deployed in medicine, engineering, and finance through the 1980s. The approach appeared capable of scaling intelligence by scaling knowledge bases and improving search heuristics.
Why it matters
How it works
Knowledge representation and inference
The building blocks of a symbolic AI system are a knowledge base and an inference engine. The knowledge base encodes facts and rules in a formal language — first-order predicate logic, a production-rule formalism, a semantic network, or a frame system. The inference engine applies rules of deduction (forward chaining from facts to conclusions, or backward chaining from goals to their prerequisites) to derive new facts or answer queries.
The power of the approach depends entirely on the quality of the knowledge base. Writing it requires knowledge engineers to interview domain experts and translate their expertise into explicit, formal rules — a process called knowledge acquisition, which turned out to be far harder than anticipated. Experts often could not articulate their own reasoning; much expertise is tacit, contextual, and difficult to decompose into discrete rules.
The brittleness problem and the frame problem
Symbolic systems are inherently brittle: they perform well within the boundaries of their encoded knowledge and fail abruptly outside them. There is no graceful degradation; there is no way for the system to recognize what it does not know or to improvise. This contrasts sharply with human expertise, which degrades gradually and produces reasonable guesses even in novel territory.
The frame problem crystallized a deeper difficulty. In a world that changes continuously, a symbolic AI system must track which facts remain true after an action and which change. Formalizing this commonsense understanding — that moving a cup leaves the table unchanged, that turning off a light does not extinguish memories of the room — requires an enormous quantity of explicit knowledge that humans handle automatically. Attempts to solve the frame problem by adding more rules only pushed it further; the common-sense knowledge base appeared to be effectively unbounded.
Where it goes next
Symbolic AI did not disappear; it transformed. Formal methods in software engineering, automated theorem provers, ontology languages for the semantic web, knowledge graphs in search engines, and planning systems in robotics all continue the symbolic tradition. The contemporary frontier is neuro-symbolic AI, which attempts to combine the sample efficiency and interpretability of symbolic representations with the robustness and generalization of neural networks. Neither paradigm alone has produced general intelligence; the working hypothesis is that integration will.