An AI winter is a sustained period of reduced funding and interest in AI research, typically following a hype cycle whose disappointment triggers funding cuts, talent attrition, and the collapse of commercial AI infrastructure. The term first appeared at the 1984 AAAI annual meeting, where Roger Schank and Marvin Minsky, both veterans of the 1970s winter, warned the business community that 1980s enthusiasm would collapse. Three years later it did. (Source: AI Winter (Wikipedia reference entry, retrieved May 2026))
The historical pattern
The history records two major winters and several smaller episodes:
| Period | Trigger | Severity |
|---|---|---|
| ~1974–1980 | Lighthill report (1973, UK); DARPA's post-Mansfield-Amendment turn to mission-oriented funding; ALPAC machine-translation report (1966); collapse of speech-understanding research at CMU | UK AI research dismantled (only Edinburgh, Essex, Sussex continued); DARPA cut $3M/year contracts |
| ~1987–2000 | Collapse of LISP machine market (1987); end of Strategic Computing Initiative (1988); expert systems judged too expensive and brittle; Japan's Fifth Generation Project missed its original goals | Billion-dollar industry collapsed; researchers shifted careers; neural-network funding was difficult through early 1980s |
Smaller episodes include the failure of machine translation (1966), the Minsky/Papert critique of perceptrons (1969), and DARPA's frustration with the Speech Understanding Research program (1971–75). (Source: AI Winter (Wikipedia reference entry, retrieved May 2026))
Recurring causes
Across both winters and the smaller episodes, several mechanisms recur:
- Promise inflation by researchers under funding pressure. Hans Moravec attributed the 1970s cuts to overpromising: "Many researchers were caught up in a web of increasing exaggeration. Their initial promises to DARPA had been much too optimistic." (Source: AI Winter (Wikipedia reference entry, retrieved May 2026))
- Underestimation of the commonsense-knowledge problem. Both ALPAC (1966) and the speech-understanding cancellation (1974) ran into the same difficulty: researchers had underestimated word-sense disambiguation. The systems worked in toy domains but failed at the boundary of unconstrained input.
- Brittleness in deployment. 1980s expert systems failed because they were expensive to maintain and broke at the edges of their rule bases. HARPY worked but only because of hard-wired knowledge that did not transfer to other signal-understanding tasks.
- Cyclical hardware bets. LISP machines collapsed in 1987 because general-purpose workstations got cheaper faster than specialized hardware could amortize. AI Chip Mania Sows Seeds of Its Own Destruction (Mackintosh, WSJ, May 16 2026) applies the parallel argument to 2026 GPU memory.
Debates and positions
The "no winter" counterargument
Historian Thomas Haigh argues there was no 1970s winter in any meaningful sense: ACM SIGART membership tripled between 1973 and mid-1978 (1,241 to 3,500), growing faster than ACM overall. In Haigh's account, funding cuts affected the major laboratories and largest projects but coincided with rapid growth in the global researcher population. Haigh's position is the principal empirical objection to treating periodic capital-market corrections as a full-field winter. (Source: AI Winter (Wikipedia reference entry, retrieved May 2026))
Current-cycle (2026) framings
The current AI boom (post-2012) has produced a sustained run of capability gains, but a contrarian set of commentators argues it shares structural features with past winters:
- Capital-cycle framings. AI Chip Mania Sows Seeds of Its Own Destruction (Mackintosh, WSJ, May 16 2026) frames current GPU/memory pricing and AI lab valuations as cycle-peak signals rather than new-paradigm baselines. The memory-chip industry's 1984, 2018, and 2022 cycle-peak P/Es (15×, 5.5×, and 9×) serve as explicit historical anchors in that argument.
- "Bubble" framings. Circular Financing in AI (Hartung) and AI Bubble Debate catalog claims that AI capex is being underwritten by booked-but-not-delivered revenue, GPU vendor financing, and accounting choices that mask economic exposure.
- Slowdown framings. Before Mythos (early 2025), there was widespread talk of "scaling stalling," cited in Ai Has Broken Containment Wong as the dominant framing in 2024–2025 before Mythos and Claude Opus 4.6 reset it.
These framings are contested. Mythos's emergence (April 2026), Anthropic's 80× Q1 revenue growth (April 2026), and the $30B to $50B Anthropic ARR trajectory (May 2026) are cited as the principal counter-evidence that the current cycle has more underlying economic substance than 1980s expert-system enthusiasm.
What would distinguish a third winter
By the historical pattern, a recognizable winter has involved all three of the following:
- A capability ceiling that disappoints stated expectations (1970s: speech recognition could not handle unconstrained input; 1980s: expert systems could not generalize past their rule bases).
- Commercial collapse of the dominant hardware substrate or business model (1987 LISP machines; in the current cycle, GPU vendor financing and hyperscaler capex amortization periods).
- Talent attrition, with researchers leaving the field rather than rotating between labs.
As of May 2026, none of the three had occurred.
Relationships
- instance-of: AI Bubble Debate
- depends-on: Scaling Laws (the technical premise current optimism rests on)
- related: AI Chip Mania Sows Seeds of Its Own Destruction (Mackintosh, WSJ, May 16 2026), Circular Financing in AI, AI Economic Primitives, AI Acceleration Paradox
See also
- AI Winter (Wikipedia reference entry, retrieved May 2026) — historical reference
- AI Chip Mania Sows Seeds of Its Own Destruction (Mackintosh, WSJ, May 16 2026), AI Bubble Debate — current-cycle framings