Expert Systems and the Second Winter

In the 1980s AI became a business: capture an expert’s know-how as if–then rules, and sell it. You will see how a rule-based system reasons, why companies and governments bet billions on the idea, and why the boom ended in a second AI winter.

Picture a computer salesman at Digital Equipment Corporation (DEC) around 1980. A customer wants a VAX minicomputer with a particular mix of memory, disk drives and terminals. Getting every cable, controller, power supply and cabinet right is fiendishly fiddly, and mistakes mean machines arriving that cannot be assembled. So DEC handed the job to a program that knew nothing about the world except several hundred rules about how its computers fit together. It worked, it saved money, and it helped set off the first commercial boom in artificial intelligence.

Knowledge is power

The 1970s had taught a hard lesson: general-purpose reasoning, as in the General Problem Solver, broke down on real problems. Edward Feigenbaum at Stanford, who had worked on DENDRAL since 1965, drew the opposite conclusion from that program’s success. What made an expert good was not clever general reasoning but a large body of specific knowledge, much of it rules of thumb. Capture that knowledge, he argued, and a fairly simple program could perform like a specialist in a narrow domain.

The result was the expert system, built from two parts. A knowledge base held rules of the form “IF these conditions hold THEN conclude this (or do this)”. An inference engine applied the rules: either forward, from known facts towards conclusions, or backward, starting from a possible conclusion and asking for the facts needed to confirm it. The people who interviewed experts and turned their judgement into rules practised a new craft, knowledge engineering.

See a rule engine think

Here is a toy configuration system in the spirit of DEC’s, cut down to four rules. The customer orders a computer with three disk drives. Step through the chain to watch the inference engine fire one rule after another.

Try the rules

A customer orders the mid-size model with 2 disk drives. Which rules fire?

Rule 1 needs more than 2 drives, so it never fires; Rules 2 and 3 depend on its conclusion. The chain never starts.

DEC launches a new drive that needs its own special cable. What happens with these four rules?

Rule-based systems know only what they are told. Outside their rules they do not fail gracefully — they are brittle. Someone has to notice and write a new rule.

Why could an expert system explain its answers, when many modern neural networks struggle to?

Every conclusion traces back to named, human-written rules. That transparency was a genuine strength.

MYCIN and XCON: the flagships

MYCIN was built at Stanford in the mid-1970s, chiefly by Edward Shortliffe for his doctoral work, with Bruce Buchanan and infectious-disease specialists. It advised on bacterial blood infections and meningitis: it asked a doctor questions, reasoned backwards through some 500–600 rules (the count grew between versions), and recommended antibiotics and doses. Because medicine is uncertain, each rule carried a certainty factor that MYCIN combined as evidence accumulated. In a 1979 blinded evaluation on meningitis cases, outside experts rated MYCIN’s treatment choices as acceptable at least as often as those of Stanford’s own specialists. Yet it was never used routinely on patients, for reasons that recur throughout AI history: questions of legal responsibility, doctors’ trust, and the practical difficulty of fitting a stand-alone terminal into hospital routines.

R1, later renamed XCON, was written by John McDermott at Carnegie Mellon from 1978 and went into use at DEC around 1980 to configure VAX orders. Its rule base grew from several hundred rules to thousands as DEC’s product range changed. Consensus It was a genuine commercial success, though the savings are only known as estimates: figures of around $25 million a year are commonly cited, and some accounts give more. Keeping XCON up to date also required a permanent team of engineers — an early sign of the hidden cost of hand-written knowledge.

Success stories like these sent companies rushing in. Through the mid-1980s large firms set up AI groups, start-ups sold expert-system “shells” (an inference engine waiting for rules), and AI became, for a few years, a hot investment.

The race with Japan

In 1982 Japan’s Ministry of International Trade and Industry (MITI) launched the Fifth Generation Computer Systems project, run by a new institute, ICOT, under Kazuhiro Fuchi. Its ten-year goal was a new kind of computer built for reasoning: massively parallel machines running logic programs, with the Prolog language at its heart, able eventually to converse and reason with knowledge. Japan had just overtaken the West in cars and consumer electronics, and the announcement caused alarm. Feigenbaum co-wrote a best-selling warning, The Fifth Generation (1983).

The responses came fast. Britain started the £350 million Alvey Programme (1983–1988), partly reversing the cuts that had followed the Lighthill Report. In the United States, a consortium of companies founded the Microelectronics and Computer Technology Corporation (MCC) in Austin, Texas, and DARPA launched the Strategic Computing Initiative (1983), which aimed at self-driving military vehicles, pilot’s assistants and battle-management systems and spent around a billion dollars over a decade. Europe set up its own ESPRIT programme.

The Fifth Generation project ended in 1992. It built working parallel inference machines and trained a generation of Japanese researchers, but it fell far short of its grand aims, and its special-purpose hardware was overtaken by ordinary computers that got faster every year.

Lisp machines, Cyc and the crash

Expert systems were mostly written in LISP, which ran slowly on ordinary hardware. Two companies spun out of the MIT AI Lab, Symbolics and Lisp Machines Inc. (LMI), sold workstations built to run it, as did Texas Instruments and Xerox. By the mid-1980s they were a flourishing niche. Then, around 1987, the market collapsed. General-purpose workstations from Sun and others, and soon desktop PCs, ran LISP well enough at a fraction of the price. LMI went bankrupt in 1987; Symbolics struggled on, filing for bankruptcy protection in 1993.

The deeper trouble lay in the expert systems themselves. Tap each weakness.

One ambitious attempt tackled brittleness head-on. In 1984 Douglas Lenat began Cyc at MCC, aiming to hand-encode the millions of common-sense facts that people take for granted, such as that things that are dropped fall. Cyc continued for decades, later through the company Cycorp, and amassed a huge knowledge base. Debated Whether it vindicated or refuted the hand-coding approach is still argued.

Other views
  • “A dead end”: critics argue that Cyc showed common sense cannot be written down rule by rule — after decades of work it never became the general reasoner once hoped for, while systems that learn from data overtook it.
  • “Underrated”: supporters reply that Cyc found commercial uses and that today’s language models still lack reliable, checkable knowledge — exactly what structured knowledge bases offer. Hybrid “neuro-symbolic” research keeps the idea alive.

The second winter

At the 1984 meeting of the American Association for Artificial Intelligence, Roger Schank and Marvin Minsky — veterans of the 1970s — warned that the enthusiasm was running ahead of reality and that an “AI winter” could follow. The phrase stuck. From about 1987 to 1993 it arrived: the Lisp-machine market collapsed, many expert-system companies folded or rebranded, DARPA’s new leadership scaled back the Strategic Computing Initiative, and the Fifth Generation project wound down. “AI” became a word to avoid in funding proposals.

This guide’s framing Yet the winter hid a quiet success. Rule-based systems did not disappear; they became ordinary software — in loan approvals, fraud checks, configuration tools and the “business rules engines” that still run inside banks and insurers today. They simply stopped being called AI. Their weaknesses, meanwhile, pointed to the next idea: systems that learn their knowledge from examples instead of having it typed in.

// Pause and reflect

  1. Expert systems could explain every conclusion; modern neural networks often cannot. How much should we value transparency against accuracy?
  2. Why do you think national governments reacted so strongly to Japan’s Fifth Generation announcement? Do you see similar dynamics in AI policy today?
  3. MYCIN performed well but was never used routinely. What does that suggest about the gap between a good demonstration and a deployed system?

Your notes are saved privately in this browser.

What happened next

While expert systems boomed and bust, a small community revived the idea that the 1970s had abandoned: neural networks that learn. Their key tool, backpropagation, made headlines in 1986. Continue with Backprop and the Connectionist Revival.

Key terms on this page

Sources & further reading