The first AI winter: when the hype froze (1974–1980)

The phrase AI winter usually names two slumps. The first, which took hold in the mid-1970s and lingered into the early 1980s, was not a single day or budget line. It was a slow loss of confidence: funders decided that grand claims about thinking machines had outrun what laboratories could deliver.

The Lighthill Report

In 1973, the British Science Research Council asked mathematician James Lighthill to review the state of AI research. His report was blunt. Much work, he argued, had been packaged as general intelligence but amounted to narrow tricks in toy domains. Progress on hard problems such as vision, language, and common-sense reasoning was disappointing relative to the publicity.

The report became a political weapon. UK funding for AI research was redirected. The message crossed the Atlantic: “general AI” was not paying off on the timetable promised.

DARPA pulls back

In the United States, DARPA had bankrolled ambitious speech-understanding and “broad AI” programmes through the late 1960s and early 1970s. By 1974–1975, reviews concluded that many projects were not meeting milestones. Budgets were cut or redirected toward more concrete military applications.

Labs that had hired for open-ended AI research shrank. Graduate students were warned that “AI” on a CV could hurt more than help.

The perceptron backlash

Parallel to government reviews, connectionist research took a hit from inside the field. Marvin Minsky and Seymour Papert’s 1969 book *Perceptrons proved important limits of single-layer networks. The mathematics was sound. The takeaway in funding committees was simpler: neural networks do not work.

That verdict was overstated. Multi-layer networks might escape the proofs, but nobody yet had a practical training method at scale. After Rosenblatt’s perceptron had made headlines in 1958, money for neural approaches dried up. Symbolic rule-based AI still dominated grant panels, even as its own grand projects stalled.

What kept going

The first winter was real but uneven. Some expert-system ideas incubated quietly. Statistical methods in speech and vision continued in smaller groups. Symbolic AI labs still produced useful tools in limited domains.

When commercial expert systems boomed in the 1980s, the field looked healthy again, until that boom collapsed into the second AI winter around 1987.

Why it matters

The first winter taught a lesson the second would repeat: benchmarks and demos are not products, and “almost human” press coverage is not a research plan. The recovery that followed did not come from louder promises. It came from narrower wins, better data, and eventually machine learning methods that could scale.

When today’s LLMs impress in conversation, the first winter is a reminder that illusion and funding cycles have shaped AI for longer than the current hype wave.