What was the second AI winter, and why did the 1980s AI boom collapse?

An "AI winter" is a stretch of years when money and enthusiasm for artificial intelligence dry up. The phrase borrows from "nuclear winter," the cold that scientists predicted would follow a nuclear war. AI has had two major winters. The first ran roughly from 1974 to 1980. The second, the subject here, ran roughly from 1987 to 1993, and it came straight after the biggest commercial boom AI had ever seen.

What makes the second winter so instructive is that it was predicted. In 1984, at the national conference of the American Association for Artificial Intelligence (AAAI) in Austin, Texas, a panel titled "The Dark Ages of AI" brought together Roger Schank, Drew McDermott, John McDermott, B. Chandrasekaran and M. Mitchell Waldrop. According to the published record, they discussed a deep unease among veteran researchers "that perhaps expectations about AI are too high, and that this will eventually result in disaster." Schank and Marvin Minsky, both survivors of the first winter, warned business leaders that enthusiasm had spiraled out of control. Within three years, the collapse began.

The boom: expert systems and Lisp machines

The boom of the early 1980s rested on expert systems. An expert system captures a specialist's know-how as hundreds or thousands of if-then rules, plus an "inference engine" that applies those rules to the facts of a case. The star example was XCON, the first commercial expert system, built at Carnegie Mellon University for Digital Equipment Corporation. It was estimated to save the company about $40 million over six years.

Companies everywhere followed. By 1985 corporations were spending more than a billion dollars a year on AI, mostly on in-house teams. A supporting industry grew up: software firms such as Teknowledge and IntelliCorp, and hardware makers such as Symbolics and LISP Machines Inc., which built Lisp machines, expensive workstations designed to run Lisp, the favorite programming language of AI researchers.

Governments joined in. In 1981 Japan committed $850 million, through its trade and industry ministry, to the Fifth Generation computer project, which aimed at machines that could converse, translate, interpret pictures and reason like people. Britain answered with the Alvey program in 1983, and the US Defense Advanced Research Projects Agency (DARPA) launched its Strategic Computing Initiative the same year.

The collapse, step by step

Schwarz's view was blunt: he dismissed expert systems as "clever programming" and believed AI was not the next big wave. A few projects survived, including DART, a battle management system that proved enormously valuable.

Why expert systems disappointed

The problem was not that expert systems never worked. Many did useful jobs. The trouble was what happened over time.

Imagine a rule-based system for approving equipment orders. It works well until the company launches a new product line. Now someone has to find the expert, interview them, write new rules, and check that the new rules do not clash with the old ones. Multiply that by thousands of rules and years of changes, and maintenance becomes a full-time burden. This is the knowledge acquisition bottleneck: experts' time is scarce, and much of what they know is hard to put into words.

Expert systems also could not learn from experience, and they were brittle: given an unusual case their rules had never anticipated, they could make absurd mistakes instead of degrading gracefully. Some researchers, such as James Hendler, argue that the deeper cause was earlier cuts to basic research, which left the field without new ideas to fix these weaknesses.

What came after

Winter did not mean AI vanished. Expert systems stayed useful in some specialized settings, and AI techniques spread quietly into larger products, often without being called AI. Rodney Brooks has described the idea that AI failed as a myth, since AI is all around us.

The years of the winter also overlapped with ideas that matter a great deal today. Neural networks were revived in the mid-1980s by work such as John Hopfield's 1982 network and the 1986 paper on backpropagation, and statistical, data-driven methods gained ground through the 1990s. Enthusiasm grew again from its low point in the early 1990s, and since about 2012 machine learning has drawn commercial interest and funding on a scale the 1980s never saw.

Could today's AI boom end in another winter?

Nobody can say for sure, and in early 2020 the question is being asked more often. Some features of the 1980s are visible again: very high expectations, heavy investment, and talk of human-level AI arriving soon.

There are also real differences. Today's machine learning systems are already built into products used by billions of people every day, such as search, translation, photo tagging and speech recognition, so they earn their keep in a way most 1980s expert systems never did. They also learn from data rather than relying only on hand-written rules, which makes them easier to update.

The safest lesson from the second winter is about promises, not technology. The 1984 panel worried that expectations had outrun what the field could deliver. If today's systems are described honestly, with their limits as well as their strengths, a correction may be milder. If they are oversold, a chill is more likely, even if the technology itself keeps improving.