An AI winter is a funding event before it is a scientific one

The name describes money, not knowledge

An AI winter is a period in which funding, hiring and institutional interest in the field collapse. It is not a period in which research stops being productive, and in both recognised episodes useful work continued throughout — often by people who had stopped using the term artificial intelligence to describe what they did.

The name itself was coined by researchers who could see one coming, borrowing from nuclear winter, which tells you something about how the field talks about its own cycles. It was applied retrospectively to earlier events and the boundaries are not precise.

Treating these as scientific failures misses the mechanism entirely. What failed was a set of promises made to people controlling budgets, and the promises were not usually made by the people doing the research.

The pattern has four steps and repeats reliably

First, a genuine technical result demonstrates something previously impossible. Second, that result is extrapolated — by press coverage, by institutions seeking funding, by consultants selling services and sometimes by researchers themselves — into a much broader claim about what will shortly be possible.

Third, funding arrives at a scale matched to the broad claim rather than the narrow result, and work expands to absorb it. Fourth, the broader claim fails to materialise on the promised schedule, the gap becomes undeniable, and funding withdraws faster than it arrived, taking with it work that was proceeding perfectly well.

Notice that nothing in this sequence requires the underlying research to be flawed. The failure is one of extrapolation, and the extrapolation is usually performed by parties with an interest in it being believed.

The first collapse followed early machine translation and a critical report

Postwar enthusiasm for automatic translation ran ahead of the linguistics, and an official evaluation in the mid nineteen sixties concluded that the results did not justify continued investment at the prevailing level. Funding in that area contracted sharply.

In the early nineteen seventies a commissioned report in the United Kingdom reached similarly deflating conclusions about artificial intelligence research more broadly, arguing that results on real-world problems had not matched the claims and that the combinatorial explosion in search would not be tamed by more computing power. Research support was substantially reduced in its wake.

The report’s technical argument was partly right and partly wrong, and it is still argued about. What is not in dispute is the effect: a research community that had grown on the basis of ambitious claims contracted rapidly when an authoritative document declared those claims unmet.

The second collapse followed a commercial boom

The nineteen eighties brought expert systems, specialised hardware built to run the languages they were written in, and a wave of corporate adoption. This was the first period in which the field made substantial money, and companies were founded, staffed and valued on the expectation that it would continue.

It did not continue in that form. Rule-based systems proved expensive to maintain, general-purpose workstations caught up with and passed the specialised machines on price and performance, and the market for a dedicated hardware category evaporated over a short period. The businesses built on it went with it.

Again the research was not refuted. Knowledge representation continued to be a serious subject and expert systems continued to run in the places where they suited the problem. What ended was the belief that this was the road to general machine intelligence, and the funding attached to that belief.

What the pattern does and does not tell you about now

It is tempting to run the template forward and conclude that a third winter is due. That is a prediction, it is not supported by the pattern alone, and there are real differences worth stating: current systems have large-scale paying users and demonstrable utility outside research settings, which was much less true in either previous cycle.

The counter-case is also real. Investment has been made on expectations of capability improvement that no one can guarantee, a substantial share of the spending is on infrastructure with a definite useful life, and the sector’s revenue relative to its capital expenditure is a matter of active argument among people who look at the numbers professionally.

The defensible position is that the historical pattern is about the relationship between promises and funding rather than about any technology, and that the relevant question is not whether the systems work but whether they work as well as the money assumes. Anyone offering a date should be treated as speculating, including the ones who are confident.