The AI winters were funding events, not scientific ones. Each was preceded by a specific promise made to a specific funder, and each was triggered by a specific document assessing whether the promise had been kept. Naming those documents turns a weather metaphor into a causal account you can check.
Where the term comes from
The phrase was coined by researchers, about their own field, before the second collapse happened. At the American Association for Artificial Intelligence annual meeting in 1984, Roger Schank and Marvin Minsky — both of whom had lived through the first cut — publicly warned that the enthusiasm around expert systems was a bubble and that a chain reaction of disappointment, funding cuts and reputational damage would follow. They borrowed the construction from nuclear winter.
That origin is worth keeping, because it undercuts the usual framing. The people who named the phenomenon were insiders predicting the next one, not historians labelling the last one. They were right within about three years.
The first winter: three documents
ALPAC, 1966
The Automatic Language Processing Advisory Committee reported to the US National Academy of Sciences in 1966 under the title Language and Machines: Computers in Translation and Linguistics. Its finding was blunt: after a decade of funded work, machine translation was slower, less accurate and more expensive than human translation, and there was no shortage of human translators to justify the effort. It recommended redirecting money to basic computational linguistics.
US government funding for machine translation was substantially cut, and stayed cut for roughly two decades. The full story of the promise ALPAC was assessing — three to five years to solved translation, made in 1954 — is on the machine translation page.
Lighthill, 1973
Sir James Lighthill, a fluid dynamicist rather than an AI researcher, was commissioned by the UK Science Research Council to survey the field. His report, Artificial Intelligence: A General Survey, appeared in 1973. It divided the subject into three categories, which he labelled A, B and C: advanced automation, building robots as a bridge between the other two, and computer-based studies of the central nervous system.
His conclusion was that A and C had produced real results but that category B — the general-purpose reasoning and robotics work that most people meant by AI — had not, and that its failure was structural rather than temporary. The mechanism he named was the combinatorial explosion: methods that worked on toy problems grew super-exponentially in cost as the problems approached realistic size, and no amount of hardware would rescue them. His summary judgement was that in no part of the field had the discoveries so far made produced the major impact that had been promised.
UK academic AI funding was concentrated into a small number of sites and cut elsewhere. The report was debated on BBC television in 1973, with Lighthill against John McCarthy, Donald Michie and Richard Gregory. It is the one AI funding review that has a broadcast record.
DARPA and speech, mid-1970s
DARPA’s Speech Understanding Research programme had run since the early 1970s with quantified targets for vocabulary size and error rate. Carnegie Mellon’s Harpy system met the vocabulary target. DARPA ended the programme anyway in the mid-1970s, on the view that the target had been met in a way that did not constitute progress towards usable speech understanding — a system tuned to a fixed vocabulary and a constrained grammar was not the thing that had been bought.
The 1969 Mansfield Amendment, which required Defense Department research funding to have a direct relationship to a military function, is the structural change behind this and behind several other cuts of the period. Undirected basic research in AI stopped being fundable from that source.
The boom in between
The thaw was commercial rather than scientific, and it was driven by one class of system: expert systems. A rule base that captured a specialist’s decision procedure could be sold to a corporation on a business case, which meant AI funding no longer had to come from a research council persuaded by a promise.
Three government programmes reacted to each other in quick succession. Japan announced the Fifth Generation Computer Systems project in 1982, a ten-year national effort around logic programming and parallel inference machines. The UK responded with the Alvey Programme from 1983; Europe with ESPRIT; the United States with the Microelectronics and Computer Technology Corporation in Austin from 1982 and, from 1983, DARPA’s Strategic Computing Initiative. Alarm about the Japanese programme is a documented motivation for the others.
Budget figures for these programmes are widely quoted and widely inconsistent, partly because they mix currencies, years and whether industrial matching funds are included. Rather than repeat a number this page cannot stand behind, note only the order of magnitude: each was a national programme in the hundreds of millions of dollars over about a decade, and none produced the general capability it was sold on.
The second winter: a hardware market collapses
The second collapse, usually dated from around 1987 to the early 1990s, began in a specific market rather than in a report — and this is the part most accounts get backwards. It was not primarily that expert systems stopped working. It was that the machines they ran on stopped making sense.
- Specialised Lisp workstations were the delivery platform. Symbolics, Lisp Machines Inc and Xerox sold expensive machines whose architecture was designed for Lisp. A serious expert system meant buying the hardware as well as the software.
- General-purpose workstations caught up. By around 1987, Sun workstations and Apple’s hardware ran competitive Lisp implementations at a fraction of the price, on machines that could also run everything else the organisation needed.
- The hardware market collapsed within about a year, taking with it the companies whose revenue funded the software ecosystem.
- The maintenance bills came due at the same time. Deployed rule bases had grown to thousands of rules, and the cost of keeping them consistent grew faster than their size. Organisations that had bought a capability discovered they had bought a permanent staffing commitment.
- DARPA reduced AI funding under new leadership in the late 1980s, on an explicit assessment that the field had not delivered against the Strategic Computing Initiative’s stated goals.
Japan’s Fifth Generation project ran its full ten years and concluded in 1992 without achieving its goals, and without the parallel logic machines it produced finding a market. By then the phrase “artificial intelligence” had become something research groups avoided in grant applications, which is the most concrete measure of a winter available: people rebranded as machine learning, knowledge-based systems, informatics, intelligent systems, decision support. Some of those names stuck permanently.
What grew during the winters
Every claim that AI research stopped is refuted by the publication record. Dated against the second winter specifically:
| During the second winter | Description |
|---|---|
| 1986 | Rumelhart, Hinton and Williams publish backpropagation in Nature, and the two Parallel Distributed Processing volumes appear. |
| 1989 | LeCun and colleagues apply convolutional networks with backpropagation to handwritten postal codes — a deployed, commercially used neural network. |
| 1988–1993 | IBM's statistical machine translation work, which replaced the entire rule-based paradigm that ALPAC had assessed. |
| 1992–1995 | Support vector machines, boosting, and the statistical learning theory that dominated the following decade. |
| 1995–1997 | Random forests take shape; Hochreiter and Schmidhuber publish LSTM in 1997, which remained the default sequence model until the transformer. |
The winters killed a research programme — hand-built symbolic systems with human-authored knowledge — while the statistical programme that replaced it grew inside the same decade under different names. Calling that a winter for AI conflates a field with one of its schools.
The pattern, stated so you can test it
Both winters follow the same four-step sequence, and stating it explicitly is more useful than any analogy about seasons:
- A demonstration succeeds on a restricted problem, under conditions that are not stated prominently.
- The demonstration is described to funders as an early version of the general capability, with a timeline.
- The general capability turns out to require something the demonstration did not have — usually orders of magnitude more data or compute, or a method that does not degrade as the problem grows.
- An assessor with no stake in the field reads the record, notes the gap between the promise and the delivery, and the money moves.
The test to apply to any current claim, therefore, is not whether the demonstration is impressive. It is whether the conditions under which the demonstration works are stated, and whether the path from those conditions to the general case is an engineering path or a hope. That question is what benchmark results are for when they are reported honestly, and what they conceal when they are not.
Top comments (0)