AI winter
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v8 ยท 4,803 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v8 ยท 4,803 words
Add missing citations, update stale details, or suggest a clearer explanation.
An AI winter is a period of reduced funding, waning public interest, and diminished research activity in the field of artificial intelligence. The term was coined in 1984 at the annual meeting of the American Association for Artificial Intelligence (AAAI), where researchers Marvin Minsky and Roger Schank warned that enthusiasm for AI had spiraled out of control [1]. The phrase draws on an analogy with "nuclear winter," evoking a cascading, self-reinforcing downturn: pessimism among researchers spreads to the media, which erodes public confidence, which triggers funding cuts, which stifles further research. The field has experienced two widely recognized winters, one running roughly from 1974 to 1980 and another from 1987 to 1993, each following a period of inflated expectations and overpromising [1][2]. Understanding these episodes is essential context for evaluating the current AI boom and the recurring question of whether a third winter may be approaching.
The phrase "AI winter" was first used publicly at the 1984 annual meeting of the American Association for Artificial Intelligence (AAAI). Two of the field's most prominent figures, Marvin Minsky and Roger Schank, convened a discussion to warn the business community that enthusiasm for AI had spiraled out of control [1][17]. Both researchers had lived through the funding drought of the 1970s and saw unmistakable parallels in the frenzied commercial investment of the early 1980s.
Minsky and Schank described a chain reaction that they feared was inevitable. First, AI companies would fail to deliver on their extravagant promises. Next, the press would turn hostile, amplifying every failure. Then government and corporate funders would slash their budgets. Finally, serious research would grind to a halt. They chose the nuclear winter metaphor deliberately: like the hypothesized climatic catastrophe following a nuclear exchange, an AI winter would be triggered by a relatively small number of failures but sustained by feedback loops that made recovery slow and painful [1]. Their prediction proved remarkably accurate. As one account of the episode put it, "three years later the billion-dollar AI industry began to collapse" [17].
To understand why the winters happened, it helps to appreciate the extraordinary optimism that preceded them. When the field of artificial intelligence was formally founded at the 1956 Dartmouth Conference, its pioneers were supremely confident. Herbert Simon predicted in 1957 that a computer would be world chess champion within ten years, and in print he forecast that "machines will be capable, within twenty years, of doing any work a man can do" [18]. Minsky was quoted in Life magazine in 1970 with an equally sweeping prediction: "In from three to eight years we will have a machine with the general intelligence of an average human being" [2][19].
Early successes seemed to justify this confidence. Programs like the Logic Theorist (1956) could prove mathematical theorems. ELIZA (1966) could simulate a Rogerian therapist convincingly enough to fool some users. The General Problem Solver (1957) could work through simple logical puzzles. Government agencies, particularly the Defense Advanced Research Projects Agency (DARPA), poured money into AI labs at MIT, Stanford, Carnegie Mellon, and Edinburgh with few strings attached [2].
But these early systems worked only in toy domains. They could solve puzzles and play simple games, yet they could not handle the complexity of real-world problems. The gap between demonstration and deployment would prove fatal to the field's credibility.
The first AI winter arrived in the mid-1970s, driven by a convergence of scathing criticism, theoretical disappointments, and government funding cuts.
In 1972, the British Science Research Council (SRC) commissioned Sir James Lighthill, the Lucasian Professor of Mathematics at Cambridge, to evaluate the state of AI research in the United Kingdom. His report, titled "Artificial Intelligence: A General Survey," was published in 1973 and delivered a devastating verdict [3].
Lighthill divided AI research into three categories: advanced automation (category A), computer-based studies of the central nervous system (category B), and bridge activities connecting the two, such as robotics and language processing (category C). While he was supportive of categories A and B in isolation, he was scathing about category C, arguing that researchers in these bridging areas had produced almost no results of practical value. His central critique targeted the problem of "combinatorial explosion": techniques that worked on small, constrained problems failed catastrophically when applied to real-world domains because the number of possible states grew exponentially with problem size [3]. His overall verdict was blunt: "In no part of the field have the discoveries made so far produced the major impact that was then promised" [3].
The Lighthill report had an immediate and severe impact. The British government effectively ended funding for AI research at most UK universities, with the notable exception of a few programs at Edinburgh and Essex. The report's influence extended beyond Britain. It reinforced growing skepticism in the United States, where similar doubts were already emerging among policymakers and funding agencies [3].
The theoretical blow to neural network research came even earlier. In 1969, Marvin Minsky and Seymour Papert published Perceptrons: An Introduction to Computational Geometry, a rigorous mathematical analysis of the single-layer perceptron, the architecture developed by Frank Rosenblatt in the late 1950s [4].
Minsky and Papert demonstrated that single-layer perceptrons could not compute certain fundamental functions, including XOR (exclusive or) and various topological predicates like connectedness. These were not obscure edge cases; they represented basic logical operations necessary for many practical tasks. The book further conjectured, based on what the authors admitted was "intuitive judgment" rather than formal proof, that multi-layer extensions of the perceptron would face similar limitations [4].
This conjecture turned out to be wrong. Multi-layer networks trained with backpropagation would later prove capable of learning XOR and far more complex functions. But the damage was done. By the time the book appeared, most researchers had already drifted away from connectionism, and Perceptrons provided an authoritative justification for abandoning the approach entirely. Funding for neural network research dried up almost completely, and the subfield entered a dormancy that would last over a decade [4].
In the United States, the most consequential blow was financial. DARPA had been the single largest funder of AI research throughout the 1960s, supporting open-ended, curiosity-driven work at leading universities. But the political environment shifted in 1969 with the passage of the Mansfield Amendment, which required DARPA to fund only "mission-oriented direct research" with a clear connection to specific military applications [5]. Pure, undirected AI research no longer qualified.
The effects were swift. DARPA tightened its requirements, demanding that researchers demonstrate near-term military utility for their work. In 1974, DARPA canceled a three-million-dollar-per-year contract with Carnegie Mellon University's Speech Understanding Research program after the project failed to meet its ambitious goals [5]. Similar cuts followed at other institutions.
The ripple effects spread through the entire research ecosystem. Universities closed or downsized AI labs. Graduate programs shrank. Promising young researchers left the field for more fundable areas of computer science. The term "artificial intelligence" itself became toxic in grant proposals, and researchers began using euphemisms like "informatics," "knowledge-based systems," or "computational intelligence" to avoid the stigma [5].
The first winter lasted roughly from 1974 to 1980. It was not a complete cessation of AI research, but rather a sharp contraction. Work continued in pockets, particularly in expert systems and knowledge representation, but at a fraction of the funding levels that had prevailed in the 1960s.
The first AI winter ended gradually, driven by a practical pivot. Rather than pursuing the grand ambition of general intelligence, researchers in the late 1970s and early 1980s turned to narrower, more achievable goals. Expert systems, programs that encoded human expertise as collections of if-then rules to solve problems in specific domains, emerged as the commercial face of AI.
The first major commercial success was XCON (also known as R1), developed by John McDermott at Carnegie Mellon University for Digital Equipment Corporation (DEC) starting in 1978. XCON configured VAX computer systems, a task that required matching customer orders against thousands of component compatibility rules. It was enormously successful: DEC reported that XCON and related systems saved the company more than $40 million per year, and by 1986 the system had processed roughly 80,000 orders with 95 to 98 percent accuracy [6][20].
XCON's success triggered a gold rush. By 1985, corporations worldwide were spending over one billion dollars per year on AI, most of it on in-house expert systems departments. Companies like Teknowledge, Intellicorp, and Applied Intelligence Systems attracted venture capital. Japan's Ministry of International Trade and Industry (MITI) launched the Fifth Generation Computer Systems (FGCS) project in 1982, a ten-year, multi-billion-dollar initiative to build massively parallel computers optimized for logic programming and AI [7]. In response, the United States launched DARPA's Strategic Computing Initiative (SCI) in 1983, allocating one billion dollars over a decade for advanced computing and AI research [8].
Specialized hardware flourished alongside the software boom. Companies like Symbolics, Lisp Machines Inc., and Texas Instruments manufactured dedicated Lisp machines, workstations optimized for running the Lisp programming language that was then standard in AI research. Symbolics, the crown jewel of this market, reported revenues of $115 million in 1986 [9].
The hype reached extraordinary levels. In 1984, the same year Minsky and Schank issued their warning at AAAI, Time magazine named the computer its "Machine of the Year." Business publications ran breathless cover stories about AI transforming every industry. Venture capital flowed freely. It looked, to many observers, like the dawn of the intelligent machine age.
The second AI winter was deeper, broader, and more damaging than the first. It struck at an industry that had grown far larger and more commercially exposed than the academic research community of the 1970s.
Expert systems, for all their early promise, had fundamental limitations that became increasingly apparent by the mid-1980s. They were expensive to build, requiring months or years of painstaking "knowledge engineering" to extract rules from human experts. They were brittle: even small changes to the problem domain could require extensive manual rule updates. They could not learn from experience or generalize beyond their programmed knowledge. And they were difficult to maintain, as rule bases grew to thousands of interacting rules that no single person could fully comprehend [6].
Corporations that had invested millions in expert systems began discovering that the maintenance costs exceeded the initial development costs. Systems that worked well in demonstrations failed in production. The "knowledge acquisition bottleneck," the difficulty of extracting and encoding expert knowledge, proved far more intractable than proponents had predicted [6].
1987 was the year the bottom fell out. Apple and IBM were producing increasingly powerful desktop computers that could run AI software adequately at a fraction of the cost of dedicated Lisp machines. Why spend $100,000 or more on a Symbolics workstation when a $5,000 desktop could do much of the same work? The market for specialized AI hardware collapsed virtually overnight. Half a billion dollars in market value vanished in a single year [9].
Symbolics, which had been the poster child of the AI hardware industry, saw its revenues plummet. The company would eventually file for bankruptcy in the 1990s. Lisp Machines Inc. and other competitors fared no better. The collapse sent shockwaves through the entire AI industry, signaling to investors and corporate executives that the AI boom had been a mirage [9].
The Fifth Generation Computer Systems project, which had been the catalyst for so much Western AI investment, quietly wound down without achieving its primary goals. Launched with great fanfare in 1982, FGCS aimed to develop computers capable of conversational language processing, image understanding, and inference on massive knowledge bases. The project produced interesting research in parallel computing and logic programming, but its ambitious AI goals proved far beyond the state of the art [7].
When the project formally ended on June 1, 1992, it had spent more than 50 billion yen, roughly $400 million at 1992 exchange rates, and was terminated without having met its goals [7]. The massively parallel inference machines it produced never found commercial application. The Prolog-based software ecosystem it fostered failed to gain adoption outside Japan. Perhaps most significantly, the project demonstrated that throwing money at AI did not guarantee results, a lesson that would echo in subsequent decades [7].
In the United States, the Strategic Computing Initiative suffered a similar fate. When Jack Schwartz took over DARPA's Information Processing Techniques Office (IPTO) in 1987, he assessed the AI components of the program and found them wanting. He cut AI funding "deeply and brutally," in the words of one historian, dismissing expert systems as mere "clever programming" rather than genuine intelligence [8].
The broader political environment reinforced these cuts. Congress ordered reductions in Pentagon research budgets in late 1985, and the Gramm-Rudman-Hollings Act of 1985 mandated automatic across-the-board spending cuts to balance the federal budget. DARPA lost $47.5 million from its budget. The Strategic Computing Initiative was "eviscerated"; its second phase plan was never released, and no annual reports or congressional testimony mentioned the program after 1988. By 1990, the AI goals of the original initiative had disappeared entirely [8].
The second AI winter was devastating for the field. The term "artificial intelligence" once again became professionally dangerous. Researchers relabeled their work as "machine learning," "data mining," "pattern recognition," or "computational intelligence" to avoid association with AI's tarnished brand. Corporate AI departments were disbanded. Startups folded. Graduate students were advised to pursue other specializations.
The psychological impact was arguably as damaging as the financial impact. An entire generation of computer scientists internalized the lesson that AI was a field of broken promises, a place where careers went to stall. This cultural stigma would linger well into the 2000s.
| Dimension | First AI Winter (1974-1980) | Second AI Winter (1987-1993) |
|---|---|---|
| Primary trigger | Academic criticism (Lighthill report, Perceptrons) | Commercial failure (expert systems, Lisp machines) |
| Duration | Approximately 6 years | Approximately 6 years |
| Geographic center | United Kingdom and United States | Global (US, Japan, Europe) |
| Key critical event | Lighthill report (1973) | Lisp machine market crash (1987) |
| Funding source affected | Government research grants (DARPA, SRC) | Both government (DARPA SCI) and private sector |
| Technology blamed | Perceptrons, general problem solvers | Expert systems, Lisp machines |
| Scale of investment lost | Tens of millions of dollars | Billions of dollars |
| International dimension | Mainly UK/US phenomenon | Japan's FGCS failure, global corporate retreat |
| What ended it | Expert systems boom, renewed DARPA interest | Statistical ML, SVMs, increased data and compute |
| Key lesson | Narrow AI can succeed where general AI cannot | Rule-based systems cannot scale; learning is essential |
Both AI winters followed a strikingly similar pattern, one that maps closely onto what Gartner would later formalize as the "hype cycle" for emerging technologies [10].
The cycle begins with a genuine breakthrough that sparks legitimate excitement. Early successes in constrained settings are extrapolated into sweeping predictions about imminent transformation. Funding floods in, attracted by the promise of revolutionary applications. Startups proliferate. Media coverage becomes breathless. Expectations soar far beyond what the underlying technology can deliver.
Then reality intervenes. Systems that worked in demonstrations fail in deployment. Promised capabilities do not materialize on schedule. Early adopters grow disillusioned. The press, having amplified the hype, now amplifies the disappointment. Funding dries up. Researchers flee to other fields. The technology enters a trough of disillusionment that can last years or even a decade.
The specific technical failures differed between the two winters, but the structural dynamics were identical:
The recovery from the second AI winter was gradual and driven by a fundamental shift in methodology. Rather than trying to encode human knowledge manually (as expert systems had done), researchers turned to statistical and probabilistic approaches that learned patterns directly from data [11].
Several developments drove this recovery:
Machine learning algorithms. New techniques emerged that were less ambitious than symbolic AI but far more practical. Support vector machines, introduced by Vladimir Vapnik in the 1990s, provided a principled framework for classification. Decision trees, random forests, and Bayesian classifiers found applications in spam filtering, fraud detection, and recommendation systems [11].
Growing data availability. The rise of the internet in the 1990s created vast quantities of digital text, images, and user behavior data. For the first time, data-hungry algorithms had enough material to learn from.
Increasing computing power. Moore's Law continued to deliver exponential improvements in processor speed and memory capacity, making it feasible to train more complex models on larger datasets.
Quiet practical successes. AI techniques began achieving real-world impact without the fanfare. IBM's Deep Blue defeated world chess champion Garry Kasparov in 1997. Speech recognition systems improved steadily. Recommendation engines at Amazon and Netflix used collaborative filtering to drive business value. These successes rebuilt confidence incrementally [2].
Importantly, many of these successes were not marketed as "AI." The stigma from the second winter was so strong that practitioners deliberately avoided the label, preferring terms like "machine learning," "data analytics," or "predictive modeling."
The event that definitively ended the long shadow of the AI winters occurred on September 30, 2012. A deep convolutional neural network called AlexNet, developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto, won the ImageNet Large Scale Visual Recognition Challenge with a top-5 error rate of 15.3%, compared to 26.2% for the runner-up [12]. The 10.8 percentage point margin was unprecedented; previous years had seen improvements measured in fractions of a percent.
AlexNet's victory was significant not just for its magnitude but for what it represented. The techniques it used, deep neural networks, backpropagation, and training on GPUs, were all decades old in principle. What had changed was the convergence of three factors: large labeled datasets (ImageNet contained over 14 million images), affordable parallel computing hardware (NVIDIA GPUs), and algorithmic refinements that made deep networks trainable in practice [12].
The floodgates opened. Within a few years, deep learning achieved breakthroughs in image recognition, speech recognition, natural language processing, game playing (AlphaGo defeated the world Go champion in 2016), protein structure prediction (AlphaFold in 2020), and generative modeling. The Transformer architecture, introduced in 2017, enabled the development of large language models that would culminate in systems like GPT-4 and ChatGPT.
By the early 2020s, AI had become the most heavily funded area of technology research and development worldwide. The term "AI" was no longer toxic; it was magnetic, attracting talent, capital, and attention at a scale that dwarfed anything in the field's history.
The extraordinary scale of the current AI boom has inevitably prompted questions about whether history is repeating itself. As of early 2026, a vigorous debate is underway about whether the AI industry is in a bubble that could burst, triggering a third winter.
The numbers involved are staggering and historically unprecedented. Aggregate annual AI investments, primarily for cloud-based mega data centers, exceeded $400 billion in 2025 and are projected to surpass $500 billion by 2026 [13]. Morgan Stanley estimates that global spending on data centers between 2025 and 2028 will total roughly $3 trillion, with half financed through private credit [13]. Major technology companies, including Microsoft, Google, Amazon, and Meta, have each committed tens of billions of dollars annually to AI infrastructure.
Critics point to a significant disconnect between spending and revenue. Hyperscalers committed nearly $400 billion in capital expenditure in 2025, while enterprise AI generated approximately $100 billion in actual revenue [14]. A report from MIT Media Lab in August 2025 stated that despite $30 to $40 billion in enterprise investment in generative AI, 95% of organizations were seeing zero return [14].
Julien Garran, a researcher at MacroStrategy Partnership, published a report in October 2025 calling the AI bubble "the biggest and most dangerous bubble the world has ever seen," estimating it to be 17 times larger than the dot-com bubble [14]. Bill Gurley, the prominent venture capitalist at Benchmark, has warned that massive AI infrastructure spending could face a painful correction by 2026, drawing parallels to the late-1990s telecom bubble [15].
In late January 2025, the successful launch of DeepSeek, a Chinese AI system that achieved competitive performance with major Western models at a fraction of the reported training cost, sent shockwaves through financial markets. On January 27, 2025, NVIDIA's shares dropped roughly 17% in a single day, erasing close to $600 billion in market value, the largest single-day market-cap loss in U.S. history at the time [14]. The event raised pointed questions about whether the massive capital expenditures on AI infrastructure were justified, or whether more efficient approaches could achieve comparable results for far less money [14].
The physical demands of AI infrastructure present another source of risk. The International Energy Agency projects that global electricity demand from data centers will more than double over the next five years. Rising energy costs could be passed on to consumers through higher electricity bills, potentially creating political backlash. Water consumption for cooling data centers has also drawn scrutiny, particularly in drought-prone regions [13].
Not everyone is convinced that a winter is coming. Several factors distinguish the current AI boom from previous episodes:
| Factor | 1980s AI Boom | Current AI Boom (2020s) |
|---|---|---|
| Revenue generation | Minimal real revenue | Hundreds of billions in actual AI revenue |
| User adoption | Confined to specialists | Hundreds of millions of consumer users |
| Technical foundation | Brittle rule-based systems | Deep learning with demonstrated generalization |
| Data availability | Limited | Essentially unlimited digital data |
| Hardware trajectory | Specialized, expensive, dead-end | GPUs on a Moore's Law-like scaling curve |
| Breadth of application | Narrow domains (expert systems) | Broad (language, vision, code, science, robotics) |
| Corporate integration | Separate "AI departments" | AI embedded across core business operations |
Proponents of continued growth argue that unlike expert systems, modern AI systems genuinely work across a wide range of tasks. Hundreds of millions of people use ChatGPT, Claude, and similar systems daily. AI-generated code, content, and analysis have become integrated into mainstream business workflows. The technology has real, measurable utility, not just demonstration-level capability.
Moreover, even if the most speculative investments prove unprofitable, the underlying technology is unlikely to be abandoned. The dot-com crash of 2000 destroyed enormous financial value, but it did not make the internet less useful. Similarly, a correction in AI valuations would not erase the genuine capabilities of deep learning systems.
The most likely scenario may be neither a full winter nor uninterrupted summer. History suggests that the AI industry could experience a significant financial correction, a "cooldown" in which unsustainable spending levels are reduced, weaker companies fail, and inflated valuations come down to earth, without a wholesale retreat from the technology itself. The distinction matters: the AI winters of the past involved not just financial losses but the near-total abandonment of entire research paradigms. The current technology base is broader, more commercially proven, and more deeply integrated into the global economy than anything that existed in the 1970s or 1980s.
Still, the historical pattern offers a genuine warning. Every previous AI boom has been accompanied by the sincere belief that "this time is different." In 1984, the expert systems industry was convinced it had found the path to practical AI. The cycle of overpromising, underdelivering, and backlash is not a relic of less sophisticated times; it is a structural feature of how emerging technologies interact with markets and public expectations.
The AI winters left a lasting imprint on the field. They taught researchers the dangers of overpromising, the importance of incremental progress, and the need to demonstrate practical value rather than theoretical potential. They also shaped the institutional culture of AI research in ways that persist today. The emphasis on benchmarks, reproducible results, and measurable performance metrics that characterizes modern machine learning is, in part, a reaction to the vague and unfalsifiable claims that contributed to previous winters.
The winters also delayed progress by years or even decades. Neural network research, effectively frozen by the aftermath of Perceptrons and the first winter, did not fully recover until the 2000s. Techniques like backpropagation, which David Rumelhart, Geoffrey Hinton, and Ronald Williams demonstrated in 1986, could arguably have been developed and applied a decade earlier if funding and interest had been sustained [16].
Perhaps the most important lesson of the AI winters is that the technology itself was never the fundamental problem. AI did not fail because the underlying ideas were wrong. It failed, temporarily, because expectations outran capability, and because the field's dependence on external funding made it vulnerable to shifts in sentiment. The ideas survived the winters and eventually flourished when computational resources, data availability, and algorithmic understanding caught up to the vision.