The AI Winter: Why AI Almost Died (Twice)

Continue Learning

• The Complete History of AI
• What is Artificial Intelligence?
• AI Glossary: Key Terms Explained
• AI Ethics for Beginners
• Open Source AI Guide

What Is an AI Winter?

The history of artificial intelligence is not a straight line of progress. It is a story punctuated by soaring ambition, crushing disappointment, and eventual revival. The term AI winter describes the periods when funding dried up, research stalled, and public enthusiasm for artificial intelligence collapsed. There were two major AI winters in the 20th century — and understanding them is essential for anyone who wants to understand where AI stands today.

Before we examine the winters themselves, it helps to understand the cycle that drives them. AI research tends to follow a pattern: a breakthrough sparks enormous excitement, governments and corporations pour money into the field, researchers make bold promises about imminent superintelligent machines, those promises fail to materialize, funding evaporates, and interest collapses. This cycle — hype, crash, recovery — has repeated itself multiple times in the history of AI, and its echoes can still be felt today.

The Dawn of AI: 1950s Optimism

To understand the winters, we must first understand the summer that preceded them. In 1956, a group of scientists convened at Dartmouth College for a workshop that would define an entire field. The attendees — including John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester — coined the term “artificial intelligence” and set an audacious agenda. They believed that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”

What followed was a decade of extraordinary optimism. Programs like the Logic Theorist and the General Problem Solver demonstrated that machines could perform symbolic reasoning. The Perceptron, developed by Frank Rosenblatt in 1957, showed that a simple neural network could learn to recognize patterns. Researchers predicted that a machine matching human intelligence was just ten to twenty years away. The U.S. government — particularly DARPA and its predecessor ARPA — poured funding into AI research, envisioning military applications, automatic translation, and thinking machines.

This period is sometimes called the Golden Age of AI. Labs at MIT, Carnegie Mellon, Stanford, and Edinburgh were buzzing with activity. The field attracted the brightest minds in mathematics, psychology, and computer science. Everything seemed possible.

The First AI Winter: 1974–1980

The first AI winter arrived quietly but inexorably in the early 1970s. Several forces converged to end the golden age. The most significant was the publication of the Lighthill Report in 1973. Commissioned by the British Science Research Council, the report by mathematician Sir James Lighthill was scathing. It argued that AI had failed to deliver on its promises in any of its major application areas: robotics, language processing, and general problem-solving. It described AI research as producing results that were impressive only on simplified toy problems and of no practical value in the real world.

The report led directly to the near-total shutdown of AI research funding in the United Kingdom. It resonated internationally. In the United States, DARPA had already begun to grow frustrated with the lack of practical results. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, a rigorous mathematical analysis that demonstrated the severe limitations of single-layer neural networks. The book cast doubt on whether neural approaches could ever produce general intelligence. Funding for neural network research collapsed almost overnight.

The technical limitations were real. The hardware of the 1970s was simply too slow and too expensive to run the kinds of programs that AI researchers imagined. Computers that cost millions of dollars could perform only a few hundred thousand operations per second. The problems AI was attempting to solve — translation, vision, reasoning — required billions of operations. The combinatorial explosion problem — the fact that the number of possible states in any non-trivial problem grows exponentially — made many AI approaches computationally intractable.

There was also a deeper conceptual problem. Early AI relied heavily on symbolic reasoning: encoding knowledge as explicit rules and symbols. This approach worked well for board games and logic puzzles but struggled with the messy, ambiguous, probabilistic nature of the real world. Translating a sentence from English to Russian is not simply a matter of substituting words; it requires deep contextual understanding that symbolic systems could not capture.

The first winter lasted roughly from 1974 to 1980. Research continued in pockets, but the field lost its prestige, its funding, and many of its best minds. Graduate students avoided AI as a career path. Journals published fewer papers. The dream of thinking machines seemed, to many, permanently deferred.

The Expert Systems Boom: 1980–1987

AI’s revival came from an unexpected direction: expert systems. Rather than attempting to create general intelligence, researchers took a narrower approach. They built systems that encoded the specialized knowledge of human experts — doctors, engineers, geologists — into rule-based programs. These systems could answer questions in narrow domains with impressive accuracy.

MYCIN, developed at Stanford in the early 1970s, could diagnose bacterial infections and recommend antibiotics with accuracy that rivaled human experts. XCON, developed at Carnegie Mellon for Digital Equipment Corporation (DEC), could configure computer systems from customer orders, saving the company millions of dollars annually. R1/XCON alone was said to save DEC $40 million per year by 1986.

The commercial success of expert systems triggered a new wave of investment. The Japanese government launched its ambitious Fifth Generation Computer Project in 1982, aiming to build massively parallel computers that could perform logical inference at enormous speeds. Alarmed by the prospect of falling behind, the United States and United Kingdom launched their own national AI initiatives. The DARPA Strategic Computing Initiative poured hundreds of millions of dollars into AI research. Private companies established AI departments. The commercial AI market, which barely existed in 1980, was worth $400 million by 1985 and predicted to reach $5 billion by 1990.

Companies like Symbolics and LISP Machines Inc. built specialized computers — LISP machines — optimized for running AI programs. The AI industry had its own hardware, its own programming languages, its own ecosystem. For a few years, it seemed like the field had finally found its footing.

You can explore more context in our guide to the history of AI and our explainer on what artificial intelligence actually is.

The Second AI Winter: 1987–1993

The second winter arrived with brutal speed. In 1987, the market for LISP machines collapsed almost overnight. Apple and IBM had introduced cheaper, faster general-purpose computers — workstations and personal computers — that outperformed the expensive specialized hardware that AI companies depended on. The LISP machine industry, which had seemed so promising, became economically unviable in a matter of months.

Expert systems also began to reveal their fundamental weaknesses. They were expensive to build and maintain. They were brittle — they performed well within their narrow domain but failed catastrophically at the edges. They could not learn from new data; every update required expensive manual intervention by knowledge engineers. And they could not explain their reasoning in ways that non-experts could easily verify or trust.

The DARPA Strategic Computing Initiative ended in disappointment. The autonomous vehicles, intelligent military pilots, and ship battle management systems it had promised failed to materialize. DARPA slashed AI funding. Japan’s Fifth Generation Computer Project, despite enormous investment, produced machines that were technically impressive but commercially useless. The project was quietly wound down in 1992.

By 1993, the AI industry had lost billions of dollars in investment. Hundreds of companies had failed. The term “artificial intelligence” became toxic — researchers began avoiding it, preferring to describe their work as “machine learning,” “knowledge engineering,” or “pattern recognition” to distance themselves from the tainted brand. This period saw one of the most comprehensive failures of a technology sector in modern history.

The failures were not merely financial. They were conceptual. The field had repeatedly promised more than it could deliver. The gap between what AI could do and what people expected it to do was enormous. And every time that gap became apparent, trust collapsed.

The Seeds of Revival: What Survived the Winters

Despite the collapses, important work continued. Geoffrey Hinton, Yann LeCun, and Yoshua Bengio persisted with neural network research through the darkest years. Hinton’s work on backpropagation — a method for training multi-layer neural networks — had been published in 1986. It was largely ignored during the second winter, but it would eventually become the foundation of the deep learning revolution.

Statistical approaches to machine learning, which treated learning as a problem of finding patterns in data rather than encoding explicit rules, gained ground in the 1990s. Support vector machines, decision trees, and probabilistic graphical models could be trained on data and made useful predictions without requiring human experts to encode their knowledge manually. These approaches were less glamorous than the AI of science fiction, but they worked.

The Internet provided an unprecedented abundance of data. As the Web grew through the late 1990s and early 2000s, researchers suddenly had access to millions of text documents, images, and other content that could be used to train machine learning models. Data had always been the limiting factor; now it was becoming abundant.

Hardware continued to improve according to Moore’s Law. By the 2000s, computers were a million times faster than they had been in the 1970s. Graphics Processing Units (GPUs), originally designed for video games, turned out to be extraordinarily well-suited to the matrix calculations at the heart of neural network training. The stage was being set for a revolution, though few recognized it at the time.

For a deep dive into how AlphaGo marked one of AI’s great turning points, read our piece on AlphaGo and the game that changed AI.

The Deep Learning Revolution: 2012 and Beyond

The modern era of AI began in 2012 with a single paper: AlexNet. A deep convolutional neural network developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto, AlexNet entered the ImageNet Large Scale Visual Recognition Challenge and destroyed the competition. It achieved a top-5 error rate of 15.3%, compared to 26.2% for the runner-up. The gap was so large that it forced the entire field of computer vision to abandon traditional approaches and embrace deep learning.

What made AlexNet possible was the confluence of three factors that had been building for decades: large datasets (ImageNet contained over a million labeled images), powerful GPUs (Krizhevsky used two NVIDIA GTX 580s), and the deep learning algorithms that Hinton and his students had been refining for years. All three elements had to come together, and in 2012, they did.

The years that followed saw breakthrough after breakthrough. In 2016, DeepMind’s AlphaGo defeated Lee Sedol, the world’s top Go player, in a match that shocked the world. In 2017, Google researchers published “Attention Is All You Need,” introducing the Transformer architecture that would eventually power ChatGPT, Claude, and every other large language model. In 2022, ChatGPT launched and became the fastest-growing consumer application in history.

Brush up on the concepts that power modern AI with our AI glossary and learn about the research powerhouse behind many breakthroughs at Google DeepMind.

Lessons from the AI Winters

The AI winters carry lessons that remain urgently relevant. The most important is the danger of hype. When researchers and companies make extravagant promises that cannot be kept, they damage trust and invite backlash. The collapse of trust is more damaging than any technical failure, because it drives away the funding, talent, and public goodwill that the field needs to progress.

The winters also illustrate the importance of fundamental research. The breakthroughs that eventually ended the second winter — backpropagation, statistical learning, transformer architectures — were developed by researchers who continued working on basic science during the lean years, often without commercial funding or institutional support. Hinton, LeCun, and Bengio won the Turing Award in 2018, decades after they began the work that made it possible.

There is also a lesson about the relationship between hardware and software. AI research has often been held back not by a lack of good ideas but by insufficient computing power to test those ideas. The deep learning revolution was possible because GPUs became powerful enough to train large neural networks in days rather than years. Future breakthroughs may similarly be unlocked by hardware advances — quantum computing, neuromorphic chips, or custom AI accelerators.

Finally, the winters remind us that intelligence — artificial or otherwise — is harder than it looks. The challenges that defeated AI researchers in the 1970s and 1980s — common sense reasoning, language understanding, robust perception — are still not fully solved today. Current AI systems are extraordinarily capable in narrow domains but remain brittle, unreliable, and often surprising in their failures. The history of AI winters is partly a history of underestimating how hard general intelligence really is.

10 Signals of a Possible AI Winter Worth Watching

• Capex outpacing revenue at the hyperscaler level. AI infrastructure spending compounding faster than monetizable revenue is the classic pre-winter pattern.
• Enterprise AI deployments below ROI expectations. Pilots that do not graduate to production. Most enterprise AI in 2026 still falls in this bucket.
• Hardware oversupply at the GPU layer. Excess GPU capacity at cloud providers without proportional demand is a leading indicator.
• Reduced VC funding for AI startups. Funding pullback follows hype peaks. Watch for the contraction signal.
• Regulatory friction that slows deployment. EU AI Act enforcement, state-level US regulation, sector-specific rules can compound to slow adoption.
• Model improvement plateauing on key benchmarks. If the model-improvement curve flattens, the perceived AGI horizon recedes; funding dynamics shift.
• Public sentiment shifts toward AI skepticism. Survey data showing eroding trust correlates with policy and capital changes.
• High-profile failures or incidents. AI mistakes that make headlines shape policy and procurement decisions disproportionately.
• Industry consolidation past the frontier-lab tier. Many AI startups acquired or shut down would signal market re-shaping.
• The narrative shifts from inevitable to overhyped. Tech press tone is a lagging indicator but worth tracking. The pendulum swing matters.

Could There Be a Third AI Winter?

With AI investment at all-time highs — hundreds of billions of dollars pouring into AI companies, chips, and data centers — the question of whether a third winter could occur is more relevant than ever. There are real risks. The current generation of large language models has significant limitations: they hallucinate facts, struggle with complex reasoning, and require enormous compute resources. If these limitations prove intractable, or if the economic returns on AI investment disappoint expectations, the conditions for another winter could develop.

There are also geopolitical risks. Export controls on AI chips, competition between the United States and China, and regulatory uncertainty could all disrupt the flow of investment and talent that sustains the current boom. History shows that AI winters can arrive quickly, often triggered by a few high-profile failures or a change in government priorities.

At the same time, today’s AI is qualitatively different from the expert systems of the 1980s. The capabilities of large language models — their ability to write code, analyze documents, explain concepts, and engage in complex reasoning — represent genuine value for businesses and individuals. The commercial adoption of AI is far deeper and more widespread than anything seen in previous cycles. This suggests that even if investment slows or a correction occurs, the field is unlikely to experience the kind of near-total collapse that characterized the first two winters.

The lesson of history, though, is humility. Every generation of AI researchers has believed it was on the verge of a permanent breakthrough. The winters happened not because the researchers were wrong that AI could be transformative, but because they underestimated how long the journey would take. The question for today’s AI leaders is whether they have learned from those who came before.

Frequently Asked Questions

What caused the first AI winter?

The first AI winter (1974–1980) was caused primarily by the Lighthill Report in the UK, which criticized AI research for failing to deliver practical results, and by the limitations of available hardware. DARPA and other government funders cut back drastically when promised results did not materialize. The publication of Perceptrons by Minsky and Papert also killed funding for neural network research specifically.

What caused the second AI winter?

The second AI winter (1987–1993) was triggered by the collapse of the LISP machine hardware market, the failure of expert systems to scale beyond narrow domains, and the cancellation of major government AI initiatives including DARPA’s Strategic Computing Initiative and Japan’s Fifth Generation Computer Project. The combination of commercial failure and unmet expectations drove a severe funding contraction.

How long did the AI winters last?

The first AI winter lasted approximately six years, from around 1974 to 1980. The second AI winter lasted approximately six years as well, from the collapse of the LISP machine market in 1987 to the gradual recovery of interest in statistical machine learning in the early 1990s. However, the full revival of neural network research did not occur until the 2010s.

What is an expert system in AI?

An expert system is a type of AI program that encodes the knowledge of human specialists — doctors, engineers, geologists — into a set of explicit rules and uses those rules to answer questions or make decisions. Expert systems were the dominant AI technology of the 1980s. They worked well in narrow domains but were expensive to build, difficult to maintain, and unable to learn from new data.

Is AI in a bubble today?

This is a subject of genuine debate among researchers and investors. Some argue that the current AI investment boom shares characteristics with previous cycles that led to winters: extravagant promises, massive capital deployment, and uncertain near-term returns. Others argue that today’s AI is fundamentally more capable and commercially valuable than previous generations, making a severe winter unlikely. History suggests that some correction is probable, but whether it will be as severe as previous winters is unknown.

Sources

This article draws on official documentation, product pages, and industry reporting. Specific sources are linked inline throughout the text.

Last reviewed: April 2026

You May Also Like

• What Is Artificial Intelligence
• Best AI Tools for Beginners
• How to Use AI
• AI Tools Directory
• Best Free AI Courses