The Second AI Winter

The Second AI Winter (roughly 1987–1993) was a period of dramatic funding cuts, broken promises, and widespread disillusionment with artificial intelligence research, following an earlier collapse in the mid-1970s. It matters because it shaped the cautious, results-driven culture of modern AI and forced researchers to abandon fragile, rule-based approaches in favor of statistical and learning-based methods that eventually powered today's revolution. The counterintuitive fact: this apparent failure era quietly incubated the neural network and machine learning ideas that would, decades later, conquer the field entirely.

The core insight

The Second AI Winter was not simply a failure of intelligence — it was a failure of a specific philosophy of intelligence. Researchers had bet everything on the idea that you could hand-craft human knowledge into rigid logical rules, and the real world turned out to be far too messy, ambiguous, and vast for that approach to scale. The "aha moment" is realizing that the winter wasn't a dead end but a natural selection event: the brittle approaches died, and the adaptive, data-hungry approaches that survived would become the foundation of modern machine learning. Without the winter, we might still be trapped in the illusion that intelligence is just a very long list of "if-then" rules.

How It Works

To understand the Second AI Winter, you need to understand what was booming just before it — and why that boom was built on sand.

The Expert Systems Boom

Through the early-to-mid 1980s, AI was riding high on the success of so-called expert systems. These were computer programs designed to mimic the decision-making of human specialists — doctors, geologists, financial advisors — by encoding their expertise as thousands of hand-written logical rules. A medical expert system might contain rules like: "IF the patient has a fever AND a rash AND has recently traveled, THEN consider tropical disease." These systems genuinely impressed people. XCON, built for Digital Equipment Corporation, saved the company an estimated $40 million per year by automatically configuring computer orders. The Japanese government announced its ambitious Fifth Generation Computer Project, promising to build thinking machines within a decade. Investment poured in. AI companies multiplied. The future seemed to be arriving on schedule.

Why the Rules Broke Down

The fatal flaw was hidden in plain sight: real-world knowledge does not fit neatly into rules. Consider trying to write every rule needed to understand a single English sentence. Words have multiple meanings. Context matters enormously. Exceptions exist to every exception. To make a medical expert system handle even a fraction of real clinical cases, engineers needed to write tens of thousands of rules — and those rules constantly contradicted, overlapped, or simply failed on cases the original experts hadn't anticipated. Maintaining these systems was a nightmare. Every time the world changed — a new drug, a new regulation, a new edge case — humans had to go back in and revise the rules by hand. The systems could not learn or adapt. They were, in essence, extremely elaborate lookup tables dressed up in the language of intelligence.

researcher · Marvin Minsky | 1927 – 2016 | Co-founder of AI and early critic of over-promising

Marvin Minsky was one of the founding figures of artificial intelligence, co-founding the MIT AI Lab and writing landmark works on perception and cognition. His early optimism about AI timelines — including a famous 1970 quote suggesting human-level AI was "from three to eight years away" — contributed to the culture of over-promise that made the eventual crash so severe. Surprisingly, Minsky spent his later career deeply skeptical of the neural network approaches that would eventually succeed, famously co-authoring a 1969 book that highlighted the limitations of early neural models in ways that dampened funding for that entire research direction for years.

The Collapse Cascades

By the late 1980s, the cracks were impossible to ignore. The Japanese Fifth Generation project failed to deliver its promised thinking machines. The American government's strategic computing initiative wound down without transformative results. Companies that had spent millions building expert systems discovered they could not maintain or scale them. The hardware that AI researchers had been selling — specialized Lisp machines built specifically to run symbolic AI programs — was suddenly being undercut by cheaper, faster general-purpose workstations. Demand evaporated almost overnight. Funding agencies, burned twice now (the First AI Winter had occurred in the mid-1970s after earlier hype collapsed), became deeply skeptical. The label "AI" became toxic in grant applications. Researchers learned to rebrand their work as "machine learning," "knowledge engineering," or "intelligent systems" simply to keep funding alive. This rebranding, born of desperation, had an unintended consequence: it forced researchers to be more rigorous, more modest, and more focused on measurable results.

researcher · Roger Schank | b. 1946 | Expert in natural language understanding and AI education

Roger Schank was a prominent AI researcher at Yale who built some of the most ambitious natural language understanding systems of the 1970s and 1980s, attempting to teach computers to read and truly comprehend stories. He became one of the most vocal critics of the expert systems boom, arguing that brittle rule-following was fundamentally the wrong approach to intelligence. Schank later left traditional academia to focus on AI-based education software, and he is notably candid in his writings about how the AI community's hunger for funding and attention led it to make claims it could not support.

Figure
The Rise and Fall of Expert Systems
AI Funding & Confidence: The Road Into Winter Investment & Hype 1980 1983 1986 1989 1992 1995 Expert Systems Peak ~1985: XCON, 5th Gen hype Lisp machine market collapses, 1987 Neural nets & ML quietly growing Second AI Winter ~1987–1993 Expert Systems hype cycle Neural net / ML research

Intuition

The Mental Model

Imagine a city that becomes famous for a miraculous new medicine. Pharmaceutical companies rush in, investors pour in money, the government announces a national program to cure every disease within ten years. The medicine really does work — but only for a narrow set of conditions under very specific circumstances. Doctors begin prescribing it for everything, and it fails spectacularly. Patients suffer, money is wasted, and the government yanks its funding. The city empties out. But in a small, unfashionable laboratory on the edge of town, a different group of researchers kept quietly working on a completely different approach — one that didn't promise miracles, that published modest results in obscure journals, and that nobody was paying much attention to. They were learning, iterating, getting a little better every year. The Second AI Winter was exactly this: the spectacular city of expert systems collapsed, and the quiet laboratory of neural networks and statistical learning survived — barely — to eventually become the foundation of everything we call AI today. The winter was devastating for careers and companies, but it was also a purifying fire that burned away the overconfident and left behind the careful.

Historical Development

Origins · 1 minhow this was discovered
History
1980
Digital Equipment Corporation deploys XCON, an expert system that configures VAX computer orders — widely celebrated as proof that commercial AI had arrived, it handles over 80,000 orders per year by 1986.
1982
Japan announces its ambitious Fifth Generation Computer Project, promising to build massively parallel thinking machines and natural language computers within ten years, triggering alarm and imitation in the United States and Europe.
1983
DARPA launches the Strategic Computing Initiative with $600 million in funding, betting on AI-powered autonomous vehicles, battle management systems, and pilot's assistant software — one of the largest AI funding programs in history to that point.
1986
Geoffrey Hinton, David Rumelhart, and Ronald Williams publish their landmark paper reintroducing backpropagation as a practical training method for neural networks — a quiet but seismic event largely ignored by the mainstream AI community still dazzled by expert systems.
1987
The market for Lisp machines — specialized hardware built to run symbolic AI programs — collapses almost overnight as cheaper general-purpose workstations from Apple and Sun Microsystems prove equally capable, wiping out companies like Symbolics and Lisp Machines Inc.
1988
The expert systems industry begins a rapid contraction; maintenance costs, brittleness, and inability to scale become impossible to hide, and major corporate clients begin quietly shelving their AI investments.
1991
Japan officially acknowledges that the Fifth Generation Computer Project has failed to meet its goals; DARPA's Strategic Computing Initiative winds down without delivering its promised autonomous military systems.
1992
The term "AI winter" is coined by researchers Eric AI Horowitz and Peter Cheeseman at a private meeting, giving a name to the chill that practitioners had been feeling for years — a moment of honest reckoning within the community.
1993
Neural network research begins attracting renewed attention as Yann LeCun's convolutional networks demonstrate practical success in reading handwritten digits for the US Postal Service — one of the first real-world machine learning deployments that would foreshadow the deep learning revolution.

Why It Matters

Real-World Impact
The Survival of Neural NetworksThe Second AI Winter forced neural network researchers like Geoffrey Hinton and Yann LeCun to operate with minimal resources and under constant skepticism, which paradoxically produced focused, rigorous work. The backpropagation methods refined during this period became the direct ancestors of every modern deep learning system, from GPT-4 to image recognition to voice assistants.
The Birth of Machine Learning as a DisciplineBecause the label "AI" became toxic during the winter, researchers rebranded their work as "machine learning" and pivoted toward statistical, data-driven methods that could demonstrate measurable, reproducible results. This cultural and terminological shift created the discipline of machine learning as distinct from classical AI — a distinction that still shapes how universities teach and companies hire today.
Realistic Benchmark CultureThe humiliation of the expert systems era baked a healthy skepticism into modern AI culture. Today's AI research community insists on standardized benchmarks, reproducible results, and careful claims — practices that emerged directly as a reaction to the unjustified promises of the 1980s boom.
The Lean Startup of AI CompaniesThe collapse of AI hardware and software companies during the winter taught the industry brutal lessons about the gap between research demonstrations and production-ready systems. Modern AI product development reflects this, with far more emphasis on deployment, scalability, and real user feedback before confident announcements.
Government Funding PhilosophyDARPA and other funding agencies emerged from the Second AI Winter with a changed philosophy, moving away from grand unified AI programs and toward funding specific, measurable capabilities. This "DARPA hard problem" approach — betting on concrete, testable goals rather than vague intelligence promises — became a model for government-funded research that persists today.
The Expert Systems Legacy in Business Rules EnginesNot everything from the expert systems era died. The core idea of encoding business logic in structured rules survived and evolved into modern business rules engines and decision management systems used in banking, insurance, and compliance — humble, honest descendants of the grand AI dreams of the 1980s.
Japan's Long AI CautionThe failure of the Fifth Generation Project left Japan's technology industry cautious about AI investment for a generation, contributing to its relative underrepresentation in the deep learning boom of the 2010s — a striking example of how a single historical failure can shape an entire nation's technological culture for decades.

Prerequisites

What This Leads To

See Also

The First AI Winter · Geoffrey Hinton and Backpropagation · Expert Systems · The History of DARPA and AI Funding · Yann LeCun and Convolutional Networks · The Symbolic vs. Connectionist Debate · The ImageNet Competition

Before this