Second AI Winter
From approximately 1987 to 1993, AI research entered a second sustained contraction as commercial expert-system vendors collapsed, the Lisp machine market failed, and DARPA substantially reduced funding for AI programmes following unmet expectations from the preceding boom.
Background
Through the late 1970s and into the 1980s, expert systems were the dominant commercial face of AI. The idea was straightforward: capture the knowledge of a human specialist as a large set of rules, then let a computer apply those rules to new problems. Digital Equipment Corporation’s R1 system, also known as XCON, showed the approach could work in practice. R1 configured VAX computer orders automatically, and by the early 1980s it was saving the company millions of dollars a year. Vendors and investors paid attention.
A hardware market grew up around this. Symbolics and Lisp Machines Inc built dedicated workstations designed to run Lisp, the programming language most AI researchers used at the time. These machines were expensive, but companies were buying them. DARPA, the Defense Advanced Research Projects Agency, poured money into the field through its Strategic Computing Programme, which began in 1983 with ambitions to produce autonomous vehicles, intelligent military interfaces and other AI-driven systems. Japan’s Ministry of International Trade and Industry launched the Fifth Generation Computer Systems project, aiming to produce logic-based machines that would leapfrog Western computing. There was genuine momentum, and genuine money behind it.
The limitation was one that boosters preferred not to discuss. Expert systems worked inside a narrow domain because a team of engineers had painstakingly written rules to cover it. Move outside that domain and the system failed immediately. Maintaining the rule base as circumstances changed was slow and expensive. There was no way to have the system learn new rules on its own. The brittleness was intrinsic, not a problem that more engineering hours would eventually fix.
What happened
Around 1987 several pressures arrived at once. Cheaper general-purpose workstations from Sun Microsystems and others began to match the performance of dedicated Lisp machines at a fraction of the price. The commercial case for buying a Symbolics machine collapsed quickly. Both Symbolics and Lisp Machines Inc ran into serious financial difficulty, and the specialist hardware market effectively disappeared. Companies that had invested in expert-system deployments found that keeping them running cost more than expected and that the systems could not be extended to cover problems adjacent to their original brief.
DARPA reviewed the Strategic Computing Programme and concluded that its AI components had not delivered what the programme’s early projections had suggested. Funding was scaled back substantially. The contraction was not a single announcement but a series of budget decisions across several years. Marvin Minsky, at the Massachusetts Institute of Technology, and Roger Schank, then at Yale University, had both warned in different ways that the field’s targets were unrealistic, though their criticisms were often directed at different aspects of the problem. The broader research community was left without the commercial revenue and government grants that had sustained large teams through the boom years.
By the early 1990s the term “AI winter,” borrowed from the “nuclear winter” hypothesis, was in common use to describe what had happened. A 1993 paper in the journal Artificial Intelligence used the phrase directly in discussing the danger of repeating the cycle. University groups absorbed researchers that commercial labs could no longer employ. Work on neural networks and statistical methods continued, mostly quietly, with far less money than the expert-system era had seen. The period lasted roughly until 1993, when the field began to stabilise on different foundations.
Why it mattered
The second AI winter demonstrated that narrowly scoped symbolic systems could not scale beyond their hand-crafted domains, discrediting expert systems as a commercial proposition and redirecting research investment toward statistical and connectionist methods over the following decade. The funding collapse forced a consolidation of AI research within universities, laying groundwork for the later machine-learning revival. It also prompted governments and funding bodies to adopt more cautious, milestone-driven evaluation criteria for AI programmes.
People
James Lighthill, Marvin Minsky, Roger Schank
Organisations
DARPA, Symbolics, Lisp Machines Inc, Fifth Generation Computer Systems Project Miti, Association for the Advancement of Artificial Intelligence
Sources
- The present danger in artificial intelligence: a plan for action.Artificial Intelligence (Elsevier).Primary source
- Strategic Computing Program, DARPA Historical Timeline.Defense Advanced Research Projects Agency (DARPA).institutional
- A history of AI winters.ACM Queue.Secondary
Cite this page
AI Achievements. (1987). Second AI Winter. Retrieved 2026-08-22, from https://achievements.ai/milestone/second-ai-winter
@misc{achievements_second_ai_winter,
title = {Second AI Winter},
author = {{AI Achievements}},
year = {1987},
url = {https://achievements.ai/milestone/second-ai-winter}
}