Overview
The Second AI Winter (roughly 1987–1993) arrived not from lack of ambition, but from the collapse of an entire industry that had overestimated how far narrow AI techniques could scale. The expert systems boom of the early 1980s ended in cascading commercial failures, abandoned government programs, and a retreat from AI that would last half a decade.
Unlike the First AI Winter — driven by theoretical disappointment — the Second was a market crash. Real money had been invested, real companies had been built, and real products had failed to deliver on their promises. The fallout was deeper and more disillusioning.
The LISP Machine Collapse
The expert systems industry had built a hardware ecosystem around it. LISP machines — specialized computers optimized for running LISP, the dominant AI programming language — became a $400 million market by 1983. Companies like Symbolics and Lisp Machines Inc. sold workstations priced at $50,000–$100,000 to corporations eager to build internal expert systems.
Then, in 1987, the market for LISP machines collapsed almost overnight. The reason: general-purpose desktop computers — Sun workstations, later PCs — became powerful enough to run LISP software at comparable speeds, at a fraction of the cost. Symbolics, once a thriving company with hundreds of employees, began a slow decline that ended in bankruptcy. The specialized hardware advantage had evaporated.
This was AI’s first experience with what would later be called commoditization of compute — a dynamic that would define the field’s economics for decades.
The Expert Systems Backlash
The deeper problem was the expert systems themselves. Companies had invested millions building and maintaining them, only to discover the knowledge acquisition bottleneck at scale:
- Rules were brittle: a system designed for one hospital’s patient population failed at another
- Maintenance was expensive: every time the domain changed, human experts had to manually update rules
- Edge cases multiplied: the real world consistently produced situations the rule-writers hadn’t anticipated
- Integration was painful: expert systems couldn’t learn from new data — they could only be updated by hand
By the late 1980s, many corporate AI departments were quietly disbanded. The promised productivity gains had not materialized. DEC’s XCON — the showcase success — still required 80 full-time employees to maintain its 10,000 rules.
DARPA’s Strategic Computing Program Ends
The U.S. Defense Advanced Research Projects Agency had launched the Strategic Computing Program in 1983, a five-year, $600 million initiative to create intelligent autonomous systems: pilot’s associates, battle management systems, autonomous land vehicles. By 1988, facing missed milestones and congressional skepticism, funding was dramatically cut.
The symbolic moment: DARPA’s Autonomous Land Vehicle project, intended to demonstrate a self-driving military truck, managed to navigate a road at 3 mph in good conditions. The gap between the program’s ambitions and its results was too wide to ignore.
The Survivors and the Underground
The Second AI Winter did not kill AI research — it drove it underground. Three communities kept working through the cold years:
Connectionists: Geoffrey Hinton, Yann LeCun, and others continued developing neural networks, especially after Rumelhart, Hinton, and Williams published the backpropagation paper in 1986. They worked largely ignored by the mainstream, at universities with modest funding.
Statisticians: Researchers began framing machine learning as a branch of statistics rather than “AI” — partly to avoid the toxicity of the AI brand, partly because statistical foundations provided mathematical rigor that symbolic AI lacked. Support Vector Machines emerged from this period.
Robotics: Rodney Brooks at MIT developed behavior-based robotics — a radically different approach that abandoned central world models in favor of distributed, reactive systems. His 1990 paper “Elephants Don’t Play Chess” challenged the symbolic AI paradigm at its foundations. His robots — simple, physical, embodied — did things the room-sized expert systems could not.
What the Winter Clarified
Looking back, the Second AI Winter was a necessary correction. It forced the field to distinguish between:
- Narrow task performance vs. genuine understanding
- Rule encoding vs. learning from data
- Brittle optimization vs. robust generalization
Max Bennett’s framework in A Brief History of Intelligence is useful here: expert systems had no mechanism for Action Simulation (modeling consequences of actions) or World Modeling (building flexible internal representations). They encoded human knowledge but couldn’t generate new knowledge — and the world is too complex to encode completely.
The researchers who survived the winter — Hinton, LeCun, Bengio — were those building systems with genuine learning mechanisms. Their time would come in 2006, when Hinton coined “deep learning” and the field began its modern renaissance.
Legacy
The Second AI Winter left lasting scars on the field’s culture:
- AI researchers learned to underpromise and overdeliver — or at least, to be more careful about what they promised
- “AI” became a toxic term in some corporate contexts; researchers rebranded as “machine learning,” “statistical learning,” or “knowledge engineering”
- The field learned that hardware + software must co-evolve: the lesson from LISP machines was that specialized hardware creates fragility
- Government and corporate funders learned to demand demonstrated results before scaling investment
The winter ended gradually. Neural networks had a quiet renaissance in the mid-1990s. By the early 2000s, with internet-scale data becoming available, the groundwork for the modern era was laid. But the scars remained — an entire generation of researchers had lived through two winters, and that institutional memory shaped how cautiously AI progress was announced for years afterward.