AI winter

What the term means

An AI winter is a stretch of time in which the money and attention going into artificial intelligence research fall sharply ^[1]. The term is usually applied to the field as a whole rather than to a single project or laboratory: work continues, but growth stops, grants become scarce, and the subject loses the public prominence it had a few years earlier.

The field has gone through several such cycles. A period of rapid progress and confident prediction is followed by disappointment and criticism, then by cuts in funding, and then, often years or decades later, by renewed interest ^[1]. Two of these downturns are conventionally dated to roughly 1974 to 1980 and 1987 to 2000 ^[1]. The dates are approximate, and different accounts place the boundaries in different years ^[1].

Before the first winter

Several separate episodes are usually listed as the run-up to the first downturn. Machine translation suffered a public failure in 1966 ^[1]. In 1969 the perceptron, an early model of a learning neuron, came in for sharp criticism ^[1]. In 1971–75 the Defense Advanced Research Projects Agency (DARPA) grew frustrated with the Speech Understanding Research programme it was paying for at Carnegie Mellon University ^[1]. In 1973 and 1974 DARPA cut back its support for academic AI research ^[1].

The funding rules behind those cuts matter. The Mansfield Amendment of 1969 required DARPA to fund mission-oriented direct research rather than basic, undirected research ^[1]. A history of US government support for computing research describes that amendment as the most important single change, because it restricted the Department of Defense to supporting basic research of direct and apparent utility to specific military functions and operations ^[2]. Before it, DARPA had followed J. C. R. Licklider's philosophy of "funding people, not projects" ^[1].

The speech programme shows how this played out. DARPA's five-year speech understanding experiment, begun in 1971, ended in 1976 with three organisations demonstrating systems, but already in 1974 the agency had cancelled a contract with the Carnegie Mellon team worth three million dollars a year ^[1]. Hans Moravec described DARPA's wording about two-million-dollar-a-year contracts being cut to almost nothing ^[1]. By 1974, funding for AI projects was hard to find ^[1].

In Britain the break came from a review rather than from a funding agency. The Lighthill Report was published early in 1973 and provoked a massive loss of confidence in AI among the academic establishment in the UK, including the funding body, a loss that persisted for almost a decade ^[3]. The report also raised questions about AI research funding in the United States and led the Department of Defense to establish a panel to assess DARPA's AI programme, known as the American Study Group ^[2].

Did it actually happen?

Not everyone accepts the winter framing. The historian Thomas Haigh has argued that there was no AI winter, because the number of researchers and the rate of discovery were speeding up rather than slowing ^[1]. Membership of the ACM's SIGART special interest group, offered as evidence that the community was not shrinking, rose from 1,241 in 1973 to 3,500 by mid-1978 ^[1].

How interest came back

In the United States the Strategic Computing Program was announced in 1983, with DARPA committing 1 billion dollars over its planned ten-year course ^[2]. In Britain the next attempt to mount a major activity in the area did not come until a Research Area Review Meeting on Intelligent Knowledge-Based Systems in September 1982, whose findings became the IKBS part of the Alvey Programme ^[3].

How the term is used

"AI winter" now works mainly as shorthand for a pattern rather than for a single event: a wave of claims outruns what the systems can do, the disappointment arrives, and the money follows the disappointment out. Because the dates are contested and because research can keep advancing while funding falls, the phrase is better read as a description of how a research field is financed and talked about than as a verdict on the technology itself.