Winters, Expert Systems, and the Cost of Overpromising Intelligence

Part of The Long Human Road to AI series.

For a few decades, the most powerful word in computer science was “intelligence.” It opened budgets, launched labs, and made front pages. It also set a trap: the more impressive the word, the more people expected a mind. When the systems on offer turned out to be useful but narrow, the disappointment had a name: AI winter.

This article is about the gap between a powerful word and a working system. It is also about what happened when researchers stopped promising general intelligence and started encoding narrow expertise. The story is not “AI failed.” It is that promises were tested against budgets, use cases, hardware, maintenance labor, and evaluation standards. Some claims collapsed. Some systems worked in narrow domains. The field kept changing under different names, methods, and institutions.

The promise meets the test

Point C1 Public evaluation reports such as ALPAC and Lighthill mattered because they tested AI-adjacent promises against measurable usefulness, not because they proved intelligence research was worthless.

In 1966, the Automatic Language Processing Advisory Committee reported to the U.S. National Research Council on machine translation. After years of optimism, the reviewers asked hard questions about translation quality, cost, and near-term usefulness. Their report, Language and Machines, became a visible warning that ambitious language-processing claims could outrun reliable performance.

In 1972–1973, Sir James Lighthill conducted a survey of artificial intelligence for the UK Science Research Council. His report criticized broad claims about general intelligence, highlighted combinatorial explosion, and argued that AI’s successes were confined to limited domains. The Lighthill report became a symbol—especially in the UK—of disappointed expectations.

Neither report was a universal verdict on all AI research. ALPAC was about machine translation and computational linguistics. Lighthill was a UK policy review with wider symbolic importance. Treating either as the single cause of a global “AI winter” would oversimplify a much messier story of budgets, institutions, and shifting confidence.

AI winter as a contested label

Point C2 The phrase “AI winter” should be handled as a contested historical label for reduced confidence, funding, and commercial enthusiasm, not as proof that research stopped.

Historians disagree about how many winters there were, when they started, and what caused them. Thomas Haigh has argued that there was no single “first AI winter” in the sense of a uniform collapse; rather, research activity continued in many areas even as public confidence cooled. Funding channels changed, some programs were cut, and the term “AI” became less fashionable in certain quarters. But laboratories, journals, and conferences did not disappear.

The useful lesson is that the health of a field cannot be read from a single headline. Confidence, money, and attention move on different schedules.

Knowledge became the center of AI

Point C3 Expert systems produced useful results in narrow domains where domain knowledge could be encoded and maintained.

By the late 1970s, Edward Feigenbaum and others argued that useful AI required domain knowledge, not abstract reasoning alone. Feigenbaum called the practice “knowledge engineering”: the craft of eliciting expertise from specialists, encoding it as rules, and building systems that could reason with them.

DENDRAL, MYCIN, and R1/XCON became the canonical examples. MYCIN, developed at Stanford, encoded infectious-disease diagnostic knowledge as a rule base with uncertainty factors and an explanation facility. R1, later called XCON, configured computer systems at Digital Equipment Corporation by applying hundreds of rules about component compatibility. These systems were not general minds. They were narrow specialists, and in their narrow territories they could match or assist human experts.

The hidden cost of expertise

Point C4 Expert-system limits included knowledge acquisition, updating, evaluation, user trust, and workflow integration, not only inference algorithms.

The rule base was only the visible part of the system. Beneath it lay the work of interviewing experts, resolving disagreements, handling exceptions, updating rules as products or diseases evolved, and explaining decisions to users who needed to trust them. The 1984 MYCIN retrospective devotes chapters to building the knowledge base, evaluating performance, designing explanations, and studying human use. The 1984 “R1 Revisited” paper describes maintenance as a continuing engineering problem, not a one-time installation.

Brittleness was a familiar symptom: a system could perform well inside its encoded boundaries and fail surprisingly outside them. The bottleneck was rarely raw computing power alone. It was the cost of keeping knowledge accurate, contextual, and aligned with real workflows.

What cooled, what continued

The contraction of the 1980s expert-system market is better described as a cooling of confidence and a shift in funding style than as a total halt. The U.S. National Research Council’s 1999 history of government support for computing research notes that AI funding changed shape through initiatives such as the Strategic Computing Program, with different expectations and accountability structures. Some work survived by being called something other than AI.

Research in machine learning, statistics, robotics, natural language processing, and computer vision continued. Many of the people and ideas that would later power data-driven AI kept working through the quieter years. The field did not stop; it reorganized.

The modern analogy

Point C5 The durable lesson for modern AI is that intelligence claims need grounded tests, maintenance plans, and institution-aware deployment criteria.

Today’s AI systems are not expert systems. They are trained on enormous datasets rather than hand-built rule bases. Yet the institutional pattern repeats: demonstrations create expectations; benchmarks discipline or inflate confidence; organizations deploy systems; and the hard questions arrive later around evaluation, maintenance, accountability, and cost. Frameworks such as the NIST AI Risk Management Framework and reports such as the Stanford HAI AI Index keep returning to test, evaluation, verification, and validation (TEVV) across the full lifecycle.

The lesson is not that rules are superior to learned models, or that hype always crashes. It is that any claim about intelligence must be paired with a plan for how it will be tested, updated, explained, and judged worth maintaining.

Further reading

For readers who want to go deeper, the primary sources behind this article include the ALPAC report, the Lighthill report, Feigenbaum’s 1977 paper on knowledge engineering, the MYCIN retrospective, the original R1 paper and its “Revisited” follow-up, and the historiographic essays by Thomas Haigh. Modern context comes from the NIST AI RMF and the Stanford HAI AI Index.

This article is part of The Long Human Road to AI. The previous article in the series is The Birth of AI; the next is Learning Machines.