An AI winter is a period of reduced funding, investment, and institutional enthusiasm for artificial intelligence (AI), typically following a phase of high expectations that available systems fail to meet. The expression describes a scientific and economic downturn rather than a complete halt in research. Histories of AI commonly distinguish two broad winters: one extending from the mid-1970s into the early 1980s, and another from the late 1980s into the early or mid-1990s. Their boundaries vary because different countries, research approaches, and industries experienced retrenchment at different times.
Meaning and scope
An AI winter involves more than disappointment with a particular product. It generally combines several developments:
- Technical shortfalls: systems perform well in demonstrations or restricted settings but fail on broader tasks.
- Revised expectations: sponsors become less confident that promised capabilities will arrive within the anticipated time or budget.
- Financial contraction: research grants, corporate programs, or purchases of AI products decline.
- Institutional consequences: companies close, research groups shrink, and researchers seek support under different disciplinary or application labels.
The term does not have a formal threshold based on investment, employment, or publication counts. It is a retrospective historical label, and its application depends partly on which institutions and technologies are being examined.
An AI winter is also distinct from the abandonment of a particular research approach. Work on artificial neural networks, for example, experienced a period of diminished visibility while other forms of AI attracted substantial support. Conversely, commercially successful applications could persist during a broader decline in enthusiasm for AI.
Background: early optimism and practical limits
AI emerged as an organized research field in the 1950s. Early programs demonstrated theorem proving, game playing, and the manipulation of symbolic expressions. These achievements encouraged expectations that general-purpose machine intelligence might be achieved relatively quickly.
Much early research used symbolic AI, representing problems through explicit symbols, rules, and search procedures. These methods could work impressively on carefully specified tasks. However, moving from simplified demonstrations to open-ended environments exposed several difficulties.
One was combinatorial explosion: the number of possibilities a program had to examine could grow rapidly with problem size. Progress in algorithms and the use of a heuristic could reduce search, but did not eliminate the underlying difficulties associated with computational complexity.
Another problem was the gap between formally described tasks and ordinary human activities. Understanding language, recognizing objects, and acting in unfamiliar situations required background knowledge and flexibility that early programs lacked. Limited processing power and memory compounded these problems, but inadequate hardware was not their only cause.
The first AI winter
Machine translation and the ALPAC report
An important precursor to the first winter was the reassessment of machine translation in the United States. Early Cold War funding supported attempts to translate languages automatically, especially Russian into English.
In 1966, the ALPAC report, produced by the Automatic Language Processing Advisory Committee, concluded that the machine-translation systems it reviewed did not justify expectations of useful, fully automatic translation in the near term. It emphasized comparison with human translation and recommended greater attention to basic language research and tools that assisted translators.
The report contributed to a substantial reduction in United States government support for machine translation. Its scope was narrower than AI as a whole, but it became an influential example of sponsors withdrawing support when demonstrations did not translate into operational value.
The Lighthill report and British funding
In 1973, mathematician James Lighthill produced the Lighthill report for the United Kingdom’s Science Research Council. It criticized much contemporary AI research, particularly the difficulty of extending results from simplified problems to realistic ones.
The report influenced a contraction and concentration of British support for AI. It did not end British AI research, and its arguments were contested by researchers. Nevertheless, it became a prominent institutional marker of the first winter.
In the United States, changes in defense research priorities also placed greater pressure on projects to demonstrate practical relevance. The first winter therefore arose from several developments rather than a single report or funding decision.
Neural-network research
The fortunes of neural-network research form a related but distinct part of this history. In Perceptrons (1969), Marvin Minsky and Seymour Papert analyzed limitations of important classes of perceptrons.
Their work is often cited as a cause of reduced interest in neural networks. That interpretation requires qualification: the results did not establish that all neural networks were incapable of solving difficult problems, and limited computing resources and the difficulty of training multilayer networks also mattered. Neural-network research continued, although it became less prominent in much of mainstream AI.
The expert-system boom
The first winter was followed by renewed investment, particularly in expert systems. These programs encoded specialist knowledge as rules and used an inference mechanism to derive conclusions. Rather than attempting general intelligence, they targeted bounded tasks such as configuring computer equipment or interpreting information within a specialized domain.
Some systems delivered useful results. Their apparent commercial potential helped create a market for AI software, consulting, and specialized computing platforms.
An expert system commonly depended on a knowledge base built through extensive work with domain specialists. This created a practical bottleneck: acquiring, checking, and updating rules was expensive. Systems could also be brittle, producing unreliable results when circumstances fell outside their encoded assumptions. Expertise that humans applied implicitly was not always easy to express as explicit rules.
The boom also supported specialized Lisp machines, designed to run software written in the Lisp programming language. These machines were closely associated with AI development during the period, making their commercial fortunes especially sensitive to changes in the AI market.
The second AI winter
The second winter developed from the late 1980s as commercial expectations surrounding expert systems and specialized AI hardware weakened.
Commercial and technical pressures
Several pressures reinforced one another:
- High development and maintenance costs: rule-based systems required substantial expert input and ongoing revision.
- Limited adaptability: systems built for one domain or organizational process were difficult to transfer to another.
- Integration difficulties: useful reasoning capabilities did not automatically produce software that fitted existing business operations.
- Hardware competition: increasingly capable general-purpose workstations reduced the advantages of specialized Lisp machines.
- An overheated market: some vendors and investors had anticipated a larger and more rapidly expanding market than materialized.
The Lisp-machine market began contracting sharply in the late 1980s. The withdrawal of vendors and investors became a visible symbol of the downturn. However, this was not a single simultaneous collapse of every AI application. Some expert systems remained in service, and rule-based techniques continued within ordinary business software.
Large research initiatives
National research programs also shaped the period. Japan’s Fifth Generation Computer Systems project, launched in 1982, sought advances in knowledge processing through approaches including logic programming and parallel inference.
The project ended in 1992 without producing the broad commercial transformation associated with its most ambitious expectations. It generated research results, but its outcome illustrated the difference between technical contributions and the realization of a sweeping industrial vision.
Elsewhere, sponsors reassessed ambitious AI programs and shifted emphasis toward narrower, demonstrable capabilities. As with the first winter, the timing and severity of retrenchment differed across institutions.
Research that continued during the winters
Neither winter amounted to the disappearance of AI. Work continued in machine learning, natural language processing, computer vision, speech recognition, and other areas.
During the 1980s, research helped establish backpropagation as a practical method for training multilayer neural networks. During the 1990s, statistical approaches became increasingly important in several AI-related fields. Researchers often emphasized measurable performance on specific tasks rather than broad claims about human-like intelligence.
Some of this activity was identified primarily with statistics, pattern recognition, optimization, or a particular application rather than with AI. Consequently, a decline in enthusiasm for the AI label did not necessarily correspond to an equivalent decline in all the research later grouped under it.
This continuity complicates simple narratives in which progress stopped during each winter and restarted only after a decisive breakthrough.
Recovery and later use of the term
Renewed growth in AI reflected a combination of improved methods, greater computing capacity, larger datasets, and commercially valuable applications. From the late 2000s and especially the 2010s, deep learning achieved major results in perception and language tasks. Graphics processing units enabled larger-scale neural-network training, while expanding digital datasets supported empirical development and evaluation.
These changes did not solve every problem associated with earlier systems. They altered which tasks could be performed effectively and which approaches attracted investment. The later expansion of generative AI again brought substantial public attention to the relationship between demonstrated capabilities, commercial expectations, and long-term promises.
“AI winter” is consequently also used in forecasts of future retrenchment. Such usage is predictive rather than a settled historical classification. A fall in investment, failure of particular businesses, or slowdown in a research approach does not by itself establish a field-wide winter.
Interpretation and limitations
The AI-winter framework is useful for examining how technical performance, public expectations, and funding decisions interact. It also has limitations.
First, it can compress geographically uneven developments into a single global chronology. A funding contraction in one country may coincide with expansion elsewhere.
Second, it can obscure differences between scientific and commercial outcomes. A research program may produce influential ideas without meeting its sponsors’ market expectations; an established application may remain useful even after investment in its sector declines.
Third, the metaphor can encourage overly simple explanations. The earlier winters were not caused solely by excessive publicity, inadequate hardware, one critical publication, or one government decision. Their causes included interacting technical limitations, procurement choices, market conditions, and institutional priorities.
Historical comparisons therefore depend on specifying what declined—funding, sales, research activity, or confidence—and which part of AI experienced that decline.