Expert System

An expert system is a computer program that imitates the decision-making of a human specialist in a narrow field – diagnosing infections, configuring computers, interpreting chemical data – by applying a large set of rules that “knowledge engineers” extract from experts through interviews. A typical rule reads: if the patient has symptoms A and B and test C is positive, then infection D is likely, with a given confidence. An “inference engine” chains such rules together to reach conclusions and can explain which rules it used. Expert systems were the most commercially successful form of symbolic AI and the basis of AI’s first commercial boom.

Classic systems

DENDRAL, begun at Stanford in the mid-1960s by Edward Feigenbaum and Joshua Lederberg with Bruce Buchanan and the chemist Carl Djerassi, inferred the structure of organic molecules from mass spectrometry data. MYCIN, developed at Stanford in the 1970s, recommended antibiotics for severe infections such as bacteremia and meningitis; in a 1979 blinded test on meningitis cases, outside specialists rated its choices at least as acceptable as those of Stanford infectious disease faculty, but it was never used in clinical practice, partly because of legal and practical concerns. Feigenbaum and Douglas Lenat later summarized the lesson as the knowledge principle: what makes such a program competent is chiefly the domain facts it holds, not the sophistication of its inference.

The boom

The breakthrough in business came with XCON (originally R1), developed by John McDermott at Carnegie Mellon from 1978 and used by Digital Equipment Corporation from 1980 to configure orders for its VAX computers; the company reported savings of about $25 million a year. By the mid-1980s many large corporations had expert system projects, startups sold “expert system shells” and specialized LISP machines, and in 1982 Japan’s government launched its ambitious Fifth Generation Computer project, prompting competing programs: MCC and the Strategic Computing Initiative in the United States, Alvey in Britain and ESPRIT in Europe.

The bust

The systems proved expensive to build and maintain. Extracting knowledge from experts was slow – the “knowledge acquisition bottleneck” – and rule bases with thousands of entries became tangled and hard to update. Systems were brittle: they failed abruptly on cases slightly outside their rules and had no common sense (see Frame Problem). Cheaper desktop computers undercut the market for LISP machines, the LISP-machine market collapsed in 1987 and many expert-system firms failed, contributing to the second AI winter.

Legacy

Rule-based systems never disappeared. They run in loan underwriting, fraud detection, tax software and clinical decision support, often under other names (see Automated Decision-Making). Their strengths – transparent reasoning that can be audited – remain attractive in regulated fields (see Explainability). Modern AI systems revive some of their ideas: companies now write down procedures for AI agents to follow (see Agent Skills), though the agents interpret those instructions with learned language understanding rather than rigid rules, and hybrid “neurosymbolic” systems pair learned models with explicit rules.

Why it matters

Expert systems showed that AI could earn money – and how quickly hype can outrun reality when knowledge has to be written by hand. Their rise and fall is a reference point whenever a new wave of AI promises to automate expertise.

Written with Claude, made by Anthropic, which created Agent Skills, a format for writing down procedures that AI agents follow, and published it as an open standard in 2025.

Related

Symbolic AI · AI Winter · Explainability · Automated Decision-Making · Frame Problem · Hype Cycle · Medical AI

Sources: Bruce G. Buchanan, Edward H. Shortliffe: Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (Addison-Wesley) (1984) · Victor L. Yu et al.: Antimicrobial Selection by a Computer: A Blinded Evaluation by Infectious Diseases Experts (JAMA) (September 21, 1979) · John McDermott: R1: A Rule-Based Configurer of Computer Systems (Artificial Intelligence) (1982) · Edward A. Feigenbaum: Knowledge Engineering: The Applied Side of Artificial Intelligence (Annals of the New York Academy of Sciences) (1984) · Douglas B. Lenat, Edward A. Feigenbaum: On the thresholds of knowledge (Artificial Intelligence) (1991) · Robert K. Lindsay et al.: DENDRAL: a case study of the first expert system for scientific hypothesis formation (Artificial Intelligence) (1993) · Anthropic: Equipping agents for the real world with Agent Skills (October 16, 2025)

Leave a Reply

Your email address will not be published. Required fields are marked *