The second AI winter: how the boom turned to bust (1987–1993)

The first AI winter (article), triggered in the 1970s by the Lighthill Report and DARPA funding cuts, had been painful but localised, mostly affecting academic labs. The second AI winter, which began around 1987 and lasted into the early 1990s, was different. This time, entire industries collapsed. Billions of dollars in private investment evaporated. The word “artificial intelligence” became toxic in boardrooms and grant applications alike.

What caused it

The winter did not have a single trigger. It was the convergence of several failures:

Expert systems hit the wall

The expert systems that had driven the 1980s boom turned out to be brittle, expensive to maintain, and impossible to scale. The knowledge acquisition bottleneck, the slow, painful process of extracting rules from human experts, proved intractable. Systems that worked in demos failed in production. Rules contradicted each other as knowledge bases grew. Updating a system with ten thousand rules was like editing a house of cards.

The Lisp machine market crashed

Throughout the boom, much of the AI industry had run on Lisp machines, specialised computers built for symbolic processing, costing between $70,000 and $150,000 each. Companies like Symbolics, once valued at over $200 million, had built their entire business around this hardware.

Then general-purpose workstations from Sun and others caught up in performance, at a fraction of the price. When expert system shells like CLIPS started running acceptably on standard hardware, the Lisp machine’s value proposition evaporated almost overnight. Symbolics’ stock collapsed from $77 to under $3. The company, along with Lisp Machines Inc. and dozens of AI startups, went bankrupt or shrank to irrelevance.

Governments pulled back

DARPA, which had poured hundreds of millions into AI through the Strategic Computing Initiative, sharply cut its AI budget in 1987. Jack Schwartz, then heading DARPA’s Information Science and Technology Office, dismissed expert systems as “clever programming” rather than genuine intelligence. Military AI projects, including components of the Strategic Defence Initiative (“Star Wars”), were cancelled when they failed to perform in realistic conditions.

Japan’s Fifth Generation Computer Systems project, which had triggered so much of the Western panic and counter-investment, quietly wound down without delivering on its promises of reasoning machines. Britain’s Alvey Programme ended similarly.

Corporate retreat

Companies that had invested millions in AI labs and expert system deployments watched returns fail to materialise. The Fortune 500 firms that had eagerly adopted the technology in the mid-1980s now walked away. Corporate AI budgets were slashed. The term “AI” itself became a liability, researchers and companies alike rebranded their work as “machine learning,” “knowledge systems,” or “advanced analytics” to avoid the stigma.

What survived

Not everything stopped. Beneath the commercial wreckage, important work continued:

• Yann LeCun demonstrated in 1989 that convolutional neural networks could be trained with backpropagation to recognise handwritten digits (LeNet). Banks later deployed it to read cheques.
• Gerald Tesauro built TD-Gammon (1992), a neural network that learned backgammon at expert level through self-play, an early triumph of reinforcement learning.
• Statistical methods in natural language processing and speech recognition quietly improved, laying groundwork for the data-driven approaches that would dominate later.

Why it matters

The second AI winter is the reason the field learned to fear hype. It demonstrated a pattern that has repeated in technology cycles: inflated expectations → over-investment → under-delivery → collapse → quiet rebuilding.

The researchers who survived the winter, by working on machine learning rather than hand-coded rules, by focusing on measurable benchmarks rather than grand claims, built the foundations for everything that followed: the statistical revolution of the 2000s, the deep learning breakthrough of the 2010s, and today’s large language models.

Every time a modern AI company makes a promise that sounds too good to be true, the second winter is the reason cautious observers ask: have we seen this before?