An AI winter is a prolonged period in which funding, commercial investment and public interest in artificial intelligence contract sharply, usually after expectations raised during a boom fail to materialize. Artificial intelligence has gone through two widely recognized winters, roughly from the mid-1970s to about 1980 and from the late 1980s to the mid-1990s, and the concept remains a core part of AI fundamentals because it explains how hype, technical limits and money interact.
The term describes an economic and institutional cycle, not a pause in scientific progress. Important work continued during both winters, but it moved into quieter niches, was often renamed, and received far smaller budgets than before.
How an AI winter develops
An AI winter develops through a recognizable boom and bust sequence: optimistic demonstrations attract money, the money creates promises, and the promises collide with hard technical limits. When sponsors see that delivered capability falls far below what was announced, they withdraw support, and the withdrawal reinforces itself as researchers, companies and students leave the field.
The phrase itself was coined in 1984 at the annual meeting of the AAAI, where researchers including Roger Schank and Marvin Minsky warned that enthusiasm for expert systems would end in disappointment. The image borrowed from “nuclear winter” suggested a chain reaction: pessimism among researchers, then in the press, then among funders, followed by cuts and the end of serious research.
The typical stages of the cycle are summarized below, with the signals that observers usually notice at each stage.
| Stage | What happens | Typical signal |
|---|---|---|
| Breakthrough demonstration | A system solves a striking but narrow task | Media coverage and bold forecasts |
| Inflated expectations | Funders and firms extrapolate from the demo | Large programs and start-up formation |
| Collision with limits | Scaling, cost or robustness problems appear | Missed milestones and quiet project cancellations |
| Contraction | Budgets are cut and companies fail | Critical official reports and layoffs |
| Rebranding and recovery | Methods survive under new names | Terms such as informatics or analytics replace AI |
The two historical AI winters
The first AI winter followed the optimism of the 1950s and 1960s, which began with the Dartmouth workshop and early programs for theorem proving and game playing. The second AI winter followed the commercial expert system boom of the early 1980s. Each had distinct triggers, but both combined a critical assessment, a funding cut and a market collapse.
Key events are listed below. Dates are approximate because a winter has no official start or end, and historians disagree about its exact boundaries.
| Period | Event | Effect on the field |
|---|---|---|
| 1966 | ALPAC report on machine translation | United States funding for translation research largely ended |
| 1969 | Perceptrons by Minsky and Papert | Single-layer neural network research lost support |
| 1969-1973 | Mansfield Amendment and DARPA restructuring | Defense money shifted to mission-oriented projects |
| 1973 | Lighthill report in the United Kingdom | Most British AI research funding was withdrawn |
| 1987 | Collapse of the LISP machine market | Specialized AI hardware vendors failed within a few years |
| Late 1980s to 1992 | Strategic Computing cuts and the end of Japan’s Fifth Generation project | Large national programs closed with modest results |
The first winter was driven mainly by government sponsors in the United States and Britain. The second was more commercial: companies had bought expert systems and dedicated workstations, then found them expensive to maintain and soon outperformed by cheaper general-purpose desktop computers.
Technical causes behind the disappointments
The deepest cause of each AI winter was a set of technical limits that were poorly understood when promises were made. Early systems worked in small, clean domains, and their builders assumed that larger problems would require only more of the same. In practice, difficulty grew far faster than problem size.
- Combinatorial explosion: search-based programs that solved toy puzzles became intractable when the number of possibilities grew exponentially.
- Commonsense knowledge gap: systems lacked the vast background knowledge that humans use without noticing, a difficulty tied to the frame problem.
- Limited computing power and data: hardware of the 1970s and 1980s was orders of magnitude too slow, and large digital datasets did not exist.
- Brittleness of expert systems: rule bases failed abruptly outside their narrow scope and required costly knowledge engineering to update.
- Moravec’s paradox: perception and motor skills proved far harder to automate than abstract reasoning.
These limits explain why symbolic AI dominated both booms and why connectionism, which needed data and compute that were not yet available, waited until the 2000s and 2010s to show its strength.
How an AI winter differs from related slowdowns
An AI winter is distinguished from ordinary fluctuations by its depth, duration and breadth: it affects public and private funding together, lasts years rather than quarters, and changes how the word AI itself is perceived. The comparison below separates it from neighboring concepts that are often confused with it.
| Concept | Scope | Typical duration | Key difference |
|---|---|---|---|
| AI winter | Whole field, public and private funding | Several years | Reputation of AI itself is damaged |
| Hype cycle trough | One technology or product category | Months to a few years | Describes adoption of a single technology |
| Market correction | Valuations of AI companies | Months | Research funding may continue unchanged |
| Research plateau | One method or benchmark | Variable | Progress stalls, but funding may not fall |
A frequent misconception is that research stopped entirely during a winter. In reality, work on machine learning, probabilistic reasoning, Bayesian networks and neural networks continued, often under labels such as pattern recognition or statistical learning, and much of it became the foundation of later booms.
Warning signs and the debate about a new winter
Analysts assess the risk of a new AI winter by comparing promised capability with delivered value. Since the deep learning revival of the 2010s and the rise of foundation models, commentators have regularly asked whether another winter is coming. There is no consensus, and forecasts in either direction should be treated with caution.
| Indicator | Healthy sign | Warning sign |
|---|---|---|
| Return on deployment | Measured savings in production | Pilots that never reach production |
| Cost per capability | Falling inference and training cost | Rising cost for marginal gains |
| Benchmark versus field results | Gains transfer to real tasks | Gains limited to test sets |
| Funding concentration | Diverse sponsors and buyers | Dependence on a few investors |
Arguments against a full winter point to broad commercial use, large revenue streams and general-purpose methods that improve with scale. Arguments for a correction point to high infrastructure costs, unreliable outputs and timelines for artificial general intelligence that may again prove optimistic. A partial cooling of investment without a collapse of research is a plausible middle outcome.
Lessons for practitioners and decision makers
The practical lesson of the AI winters is to tie claims to measurable results. Organizations that describe what a system does today, publish its limitations and fund maintenance realistically are far less exposed when sentiment turns.
- Separate demonstrations from deployable products and budget for the gap between them.
- Track total cost of ownership, including data upkeep and model maintenance, not only build cost.
- Avoid dependence on one technique, vendor or funding source.
- Communicate uncertainty in timelines instead of fixed dates for capabilities not yet shown.
In everyday language
An AI winter is a long cold spell for artificial intelligence. People get very excited, money pours in, and experts promise thinking machines soon. When the machines turn out to be much less capable than promised, investors and governments lose patience and stop paying.
The key point is simple: the winter is caused by the gap between promises and results, not by a lack of good ideas. Many useful ideas survive the cold and grow again when computers become faster and data becomes plentiful.
Seen in the real world
In the early 1980s, a manufacturer of industrial equipment might have bought an expert system to configure orders, along with a dedicated LISP workstation costing tens of thousands of dollars. The system worked well for the first product line, with a few thousand rules written by engineers and specialists.
Within a few years, every new product required new rules, the rule base grew hard to verify, and the specialist workstation was outpaced by ordinary desktop computers. Repeated across many firms, this pattern meant budgets were cut, vendors closed and AI projects were quietly abandoned, a typical picture of the second AI winter at the level of a single company.