An AI winter is a period of reduced funding, interest, and research activity in Artificial intelligence following a wave of inflated expectations that the technology of the time failed to meet. The term, coined by analogy with "nuclear winter," describes cyclical boom-bust patterns that have recurred since the field's founding at the 1956 Dartmouth workshop. Historians generally identify two major AI winters, in the mid-1970s and from the late 1980s into the early 1990s, though smaller regional slowdowns also occurred.
Each winter followed a similar pattern: early demonstrations generated media hype and heavy government or corporate investment, the underlying approach then hit fundamental limits its proponents had underestimated, funders lost patience, and research budgets collapsed, pushing many researchers to rebrand their work outside the "AI" label entirely.
First AI winter (mid-1970s)
The first winter was triggered partly by the UK's 1973 Lighthill report, commissioned by the British government, which concluded that AI research had failed to deliver on its early promises and led to major cuts in university funding across the country. In the United States, DARPA scaled back open-ended AI research funding after speech-recognition and machine translation projects underdelivered; machine translation in particular had already been damaged years earlier by a 1966 US government report finding it unable to match human quality. A compounding factor was Marvin Minsky and Seymour Papert's 1969 book demonstrating the mathematical limits of the single-layer Perceptron, which discouraged neural network research for over a decade.
Second AI winter (late 1980s to early 1990s)
The second winter followed the collapse of the commercial market for expert systems and specialized LISP machines, after both proved expensive to maintain and unable to scale to messy real-world knowledge outside their narrow domains. Japan's Fifth Generation Computer Project, a large government bet launched in 1982 on symbolic AI and logic programming, failed to produce a commercial breakthrough by its conclusion in 1992, further souring investor sentiment worldwide. The label "AI" itself became commercially unfashionable during this period, and many researchers described their work instead as "machine learning," "informatics," or "pattern recognition."
Legacy and modern debate
Recovery from each winter came from narrower, provably useful techniques rather than renewed promises of general intelligence: statistical Machine learning through the 1990s and deep learning after 2012. Since the launch of ChatGPT in 2022, commentators have periodically asked whether the current AI boom, with its enormous capital expenditure and disputed scaling laws returns, could end in a third winter. Others argue that today's AI, unlike earlier eras, already generates substantial commercial revenue and mass adoption, making a symmetrical collapse less likely even if a correction in valuations eventually occurs.