AI winter refers to periods of reduced funding, interest, and progress in artificial intelligence research, typically following inflated expectations and subsequent disillusionment. These downturns, occurring primarily in the mid-1970s and late 1980s to early 1990s, were characterized by government and corporate funding cuts, a collapse of AI startups, and a shift in focus toward narrower, more practical applications. The term draws an analogy to a "nuclear winter," emphasizing the severe slowdown of research activity.
1 Historical background
1.1 Early optimism and the Dartmouth Conference
The field of artificial intelligence was formally launched at the Dartmouth Conference in 1956, where leading researchers gathered to explore the possibility of creating machines that could simulate human intelligence. The early years were marked by extraordinary optimism: many believed that general-purpose AI—machines capable of any intellectual task—was only decades away. Early successes in domains such as theorem proving and checker-playing fueled this confidence, leading to generous funding from government agencies like the U.S. Department of Defense and the Advanced Research Projects Agency (ARPA).
1.2 First AI winter (1974–1980)
1.2.1 Lighthill Report (1973)
In 1973, the British government commissioned mathematician James Lighthill to evaluate the state of AI research in the United Kingdom. His report was highly critical, concluding that most AI work had failed to deliver on its grand promises, especially in areas like robotics and natural language understanding. The Lighthill Report led to a sharp reduction in government funding for AI in Britain and influenced similar reviews elsewhere.
1.2.2 DARPA funding cuts
In the United States, the Defense Advanced Research Projects Agency (DARPA) had been a major sponsor of AI research since the 1960s. By the early 1970s, DARPA became disillusioned with the slow progress toward practical military applications, particularly in speech recognition and autonomous vehicles. In 1974, DARPA significantly cut its AI budget, slashing funding for many university labs and causing a cascade of project terminations. This, combined with the British pullback, triggered the first AI winter.
1.3 Second AI winter (1987–1993)
1.3.1 Collapse of the Lisp machine market
During the 1980s, a resurgence of AI interest was driven by the commercial success of expert systems. Specialized hardware—Lisp machines built by companies like Symbolics, Lisp Machines Inc., and Texas Instruments—became the platform of choice for AI development. However, by the late 1980s, general-purpose workstations from Sun Microsystems and other vendors matched or exceeded Lisp machine performance at a fraction of the cost. The market for expensive Lisp hardware collapsed, bankrupting several companies and eroding investor confidence in AI.
1.3.2 Japanese Fifth Generation project overreach
In the early 1980s, Japan launched the Fifth Generation Computer Systems project, an ambitious initiative to create advanced AI machines using logic programming. The project attracted worldwide attention and prompted massive government and corporate investments in other countries, notably the UK’s Alvey programme and the US’s Strategic Computing Initiative. By the early 1990s, the Fifth Generation project had failed to meet its revolutionary goals, producing only limited results. This widely publicized failure deepened the second AI winter, as governments scaled back long-term AI funding.
2 Causes
2.1 Overhyped expectations
2.1.1 Promises of general intelligence
From the 1950s through the 1980s, AI researchers frequently promised that machines exhibiting human-level general intelligence were imminent. Such claims, repeated in academic papers and media interviews, created unrealistic public and governmental expectations. When these timelines repeatedly slipped, the resulting disappointment was intense.
2.1.2 Unfulfilled demonstrations
Early AI systems, such as ELIZA and SHRDLU, showed impressive behavior in narrow contexts, leading many to believe that broader capabilities were just around the corner. However, attempts to extend these demonstrations to real-world complexity failed. For example, machine translation projects in the 1960s produced wildly inaccurate results, and early autonomous vehicle prototypes could not navigate even simple roads. These high-profile failures were seized upon by critics and funding agencies.
2.2 Technical limitations
2.2.1 Lack of computational power
The hardware of the 1970s and 1980s was far too limited to support the combinatorial searches and massive datasets required for many AI approaches. Symbolic reasoning systems quickly ran up against memory and speed constraints, making even moderate-sized problems intractable.
2.2.2 Intractability of symbolic reasoning
Many early AI systems relied on symbolic logic and rule-based manipulation. As tasks grew in complexity, the number of rules and possible inference paths exploded. Researchers discovered that many of the problems they hoped to solve were NP-hard or undecidable; no amount of clever engineering could overcome fundamental computational barriers. This realization undercut the dominant paradigm of symbolic AI.
2.3 Funding and policy shifts
2.3.1 Government review panels
Beyond the Lighthill Report, several other government-commissioned reviews in the US and Europe delivered negative assessments of AI progress. These reviews, combined with changing political priorities, led to systematic redirection of funding away from open-ended AI research toward more immediately applicable technologies.
2.3.2 Corporate retreat from long-term research
During the second AI winter, large corporations such as IBM, Digital Equipment Corporation, and Xerox closed or downsized their in-house AI research labs. The failure of AI startups to produce sustainable revenue convinced many corporate boards that AI was a dead end, prompting a shift toward short-term product development and the abandonment of blue-sky research.
3 Consequences
3.1 Research restructuring
3.1.1 Rise of expert systems
Despite the winters, one AI subfield—expert systems—survived and even thrived during the late 1970s and 1980s. These rule-based systems encoded human expertise for specific domains (e.g., medical diagnosis, mineral exploration) and delivered tangible business value. The success of expert systems provided a pragmatic counterexample and kept AI research alive in industry.
3.1.2 Emphasis on machine learning subfields
After the second winter, many researchers turned away from symbolic, knowledge-driven AI toward statistical and data-driven approaches. Subfields such as neural networks, decision trees, and Bayesian inference gained traction because they could make progress with limited computational resources and required less handcrafted knowledge. This shift laid the groundwork for the eventual recovery.
3.2 Public perception and media narrative
3.2.1 "AI is dead" sentiment
During both winters, major newspapers and magazines published articles declaring that artificial intelligence had failed and was effectively dead. This narrative persisted for years, discouraging students from entering the field and making it difficult to attract funding even for modest projects.
3.2.2 Aftermath in academic departments
Universities responded to the funding crunch by renaming AI departments and programs, often subsuming them under broader headings such as "cognitive science," "pattern recognition," or "information processing." The word "artificial intelligence" itself became stigmatized in some academic circles, and many researchers adopted alternative labels (e.g., "machine learning," "knowledge systems") to avoid negative associations.
4 Legacy and recovery
4.1 Lessons for the field
4.1.1 Importance of incremental progress
The AI winters taught researchers the value of modest, step-by-step advancements rather than grand promises. The recovery era emphasized building on proven techniques, benchmarking performance, and publishing reproducible results—all practices that helped restore credibility.
4.1.2 Avoiding hype cycles
The field developed a heightened awareness of the dangers of overhyping results. Research communities became more cautious about media claims and established norms for tempering predictions. While hype has not entirely disappeared (as seen in later booms), the memory of the winters serves as a cautionary tale.
4.2 Revival factors
4.2.1 Increased data availability (1990s)
The rise of the internet and digitization of information in the 1990s provided AI researchers with vast amounts of training data. Machine learning algorithms, especially those based on statistical methods, thrived on this data, enabling tangible progress in tasks like text classification, speech recognition, and computer vision.
4.2.2 Rise of statistical learning methods
The 1990s saw the maturation of statistical learning theory, including support vector machines, graphical models, and ensemble methods. These techniques offered robust performance with mathematical foundations, attracting new researchers and funding. They became the backbone of practical AI systems long before the deep learning era.
4.2.3 Deep learning breakthroughs (2000s onward)
Beginning in the mid-2000s, advances in neural network training—fueled by faster GPUs, larger datasets, and algorithmic innovations like backpropagation and dropout—led to dramatic improvements in image classification, speech recognition, and natural language processing. Landmark results such as AlexNet (2012) captured renewed public attention and massive investment. The deep learning revolution marked the definitive end of the AI winter, establishing AI as a central technology of the 21st century.