Wikipedia Deep Dive

AI winter

14 min read

The Prophets of Doom Were Right

In 1984, two of the most famous artificial intelligence researchers in the world stood before a room full of their peers and delivered a warning. Roger Schank and Marvin Minsky had lived through the collapse of AI funding in the 1970s. Now they watched history repeat itself. The hype had spiraled out of control again, they said, and disappointment would certainly follow.

They called it an "AI winter."

The metaphor was deliberately apocalyptic. Like nuclear winter—the theoretical aftermath of atomic warfare where dust and debris block the sun and temperatures plunge—an AI winter would trigger a chain reaction of despair. First, pessimism would spread among researchers. Then journalists would turn skeptical. Funding agencies would slash their budgets. Serious research would grind to a halt. Careers would end. Years, perhaps decades, of progress would be lost.

Three years after that warning, the billion-dollar AI industry collapsed.

A Pattern Written in Failed Promises

The history of artificial intelligence is not a steady march of progress. It's a boom-and-bust cycle as predictable as any in economics—and just as devastating when the bust arrives. Periods of wild optimism and generous funding give way to disappointment, criticism, and years of neglect. Then, eventually, interest revives, and the cycle begins again.

Two major winters dominated the field. The first lasted roughly from 1974 to 1980. The second, longer and more severe, stretched from 1987 to 2000. But these were just the largest downturns. Smaller freezes struck repeatedly, each one triggered by a specific failure that shattered faith in AI's promises.

In 1966, machine translation collapsed. In 1969, a devastating critique of neural networks stopped that line of research cold. In the early 1970s, the American defense agency DARPA grew frustrated with speech recognition research. That same period saw British AI research dismantled almost entirely. The late 1980s brought the crash of specialized AI hardware. The 1990s saw the abandonment of expert systems that corporations had spent fortunes building.

Each failure followed the same pattern: grandiose promises, inadequate results, backlash, funding cuts.

The Translation Debacle

The story of machine translation illustrates how AI winters begin. In 1954, IBM and Georgetown University staged a public demonstration. A computer translated Russian sentences into English. Newspapers went wild. "The bilingual machine!" headlines proclaimed. "Robot brain translates Russian into King's English!" "Polyglot brainchild!"

The demonstration was, charitably speaking, a show.

The machine could handle exactly 49 carefully selected sentences using a vocabulary of just 250 words. To understand how limited this was, consider that humans need around 8,000 to 9,000 word families to comprehend written text with reasonable accuracy. The Georgetown-IBM machine could handle roughly 3% of that.

But context matters. This was the Cold War. The American government desperately wanted to read Russian scientific papers and intelligence documents. The Central Intelligence Agency believed machine translation could revolutionize espionage. Federal money poured into research programs.

Researchers were optimistic. Noam Chomsky's revolutionary work on grammar seemed to promise that language followed logical rules computers could master. Breakthroughs felt imminent.

They weren't. The researchers had underestimated something fundamental: to translate a sentence correctly, you have to understand what it means. Consider the phrase "the spirit is willing but the flesh is weak." Legend has it that early machine translation systems, passing this through Russian and back, produced "the vodka is good but the meat is rotten."

Apocryphal or not, the story captures the problem perfectly. Machines didn't know that "spirit" in this context meant willpower, not alcohol. They didn't understand "flesh" meant the human body, not butchered meat. This became known as the "commonsense knowledge problem"—machines lacked the vast background understanding humans absorb simply by existing in the world.

By 1964, the National Research Council had grown concerned enough to convene the Automatic Language Processing Advisory Committee, or ALPAC. Their 1966 report was brutal: machine translation was more expensive, less accurate, and slower than human translation. After spending twenty million dollars—a substantial sum in 1960s money—the government ended all support. Research programs shut down. Careers were destroyed.

The Death and Resurrection of Neural Networks

Today, neural networks power everything from voice assistants to image generators to language models. It's strange to remember that the field was effectively dead for over a decade.

The basic idea of neural networks—computing systems loosely modeled on how neurons connect in biological brains—emerged in the 1940s and 1950s. Early systems like the perceptron, invented by Frank Rosenblatt, generated excitement. Rosenblatt himself predicted that perceptrons "may eventually be able to learn, make decisions, and translate languages."

Sound familiar? The cycle was already beginning.

In 1969, Marvin Minsky and Seymour Papert published a book called Perceptrons. It was a rigorous mathematical analysis demonstrating the limits of what single-layer neural networks could compute. They couldn't solve certain problems that seemed trivially easy—like recognizing when two points in an image were connected.

Now, researchers already knew that multilayer networks—perceptrons stacked in layers—could solve these problems. But nobody knew how to train a multilayer network. The mathematical technique needed, called backpropagation, wouldn't be developed for years. Without a way to make these more powerful networks learn, the limitations Minsky and Papert identified seemed fundamental.

Funding for neural network research dried up almost completely through the 1970s and early 1980s. Important theoretical work continued on shoestring budgets, driven by researchers who believed in the approach even when almost no one would pay for it.

The winter thawed in the mid-1980s when John Hopfield, David Rumelhart, and others developed techniques that made neural networks practical again. Backpropagation finally provided a way to train multilayer networks. Interest revived. The field that had been abandoned as a dead end became, eventually, the foundation of modern AI.

Frank Rosenblatt didn't live to see it. He died in a boating accident in 1971, shortly after Perceptrons was published, his optimistic predictions still unfulfilled.

The Lighthill Report: How Britain Killed Its AI Research

In 1973, the British Parliament asked Professor Sir James Lighthill to evaluate the state of artificial intelligence research in the United Kingdom. The resulting document, known as the Lighthill report, became infamous in AI circles as one of the most damaging critiques the field has ever received.

Lighthill concluded that AI had utterly failed to achieve its "grandiose objectives." Nothing being done in AI, he argued, couldn't be done better in other fields of science. He highlighted a fundamental problem called "combinatorial explosion"—or, more technically, "intractability."

Here's what this means. Many AI algorithms work by searching through possible solutions. For simple problems, this works fine. But as problems get more complex, the number of possibilities explodes exponentially. An algorithm that solves a toy problem in seconds might need centuries to solve the real-world version. Lighthill argued that AI's successes were confined to carefully constructed demonstrations, while real-world problems remained utterly out of reach.

The report sparked a televised debate on the BBC. Lighthill faced off against Donald Michie, John McCarthy—one of the founders of the field who had coined the term "artificial intelligence"—and Richard Gregory. McCarthy later acknowledged that the combinatorial explosion problem had been recognized from AI's earliest days. But recognition hadn't meant solution.

The report's consequences were devastating. The British government essentially dismantled AI research across the country. Only a few universities—Edinburgh, Essex, and Sussex—continued working in the field. A decade of potential progress was lost.

Research wouldn't revive at scale until 1983, when the British government's Alvey program began funding AI again. The catalyst? Japan had announced an ambitious Fifth Generation Computer project, and Western governments panicked about falling behind. Suddenly, funding AI research seemed like a matter of national competitiveness.

The Pentagon's Frustration

In the 1960s, the Defense Advanced Research Projects Agency—DARPA, the American military's research arm—approached AI with an almost romantic attitude. The founding director of DARPA's computing division, J.C.R. Licklider, believed in "funding people, not projects." AI luminaries like Marvin Minsky, John McCarthy, Herbert Simon, and Allen Newell received millions of dollars to spend essentially however they liked. Pure research flourished.

This golden age ended in 1969 with the Mansfield Amendment, a congressional law requiring DARPA to fund "mission-oriented direct research, rather than basic undirected research." The age of exploration was over. Now researchers had to demonstrate their work would produce useful military technology—soon.

AI proposals suddenly faced harsh scrutiny. The Lighthill report from Britain and DARPA's own internal studies suggested most AI research was unlikely to produce anything practical in the foreseeable future. Money was redirected toward specific military goals: autonomous tanks, battle management systems. By 1974, funding for general AI projects had nearly vanished.

AI researcher Hans Moravec later described what went wrong in blunt terms: "Many researchers were caught up in a web of increasing exaggeration." Their initial promises to DARPA had been wildly optimistic. When they delivered less than promised, they couldn't bring themselves to promise less the next time—so they promised even more. Expectations spiraled beyond any possibility of fulfillment.

The backlash was harsh. Moravec reported that some at DARPA felt researchers needed to be "taught a lesson." Contracts worth millions of dollars annually were slashed to almost nothing.

There's an irony here. The autonomous tank project, which DARPA had prioritized, failed. But the battle management system—a tool called the Dynamic Analysis and Replanning Tool—succeeded beyond anyone's expectations. During the first Gulf War in 1991, it saved billions of dollars, repaying every cent DARPA had ever invested in AI research. The pragmatic, goal-oriented approach that had seemed so hostile to pure research ended up producing one of AI's greatest practical successes.

The Speech Recognition Disappointment

In 1971, DARPA launched an ambitious five-year experiment in speech understanding. The goal was to create systems that could recognize spoken English from a limited vocabulary in near-real time—imagine a pilot issuing voice commands to a computer.

By 1976, three organizations demonstrated working systems. Carnegie Mellon University showed two systems, called HEARSAY-II and HARPY. Bolt, Beranek and Newman, and a joint team from the System Development Corporation and Stanford Research Institute, presented their own approaches.

HARPY came closest to meeting the original goals. It worked—but only through a trick that severely limited its usefulness. The system achieved its high performance by "hard-wiring" information about expected utterances into its knowledge base. It could recognize speech, but only if speakers used words in particular orders the system expected. This wasn't the flexible voice-command system DARPA had envisioned.

DARPA felt betrayed. They believed they had been promised something transformative and received something severely limited. In 1974, they cancelled a three-million-dollar-per-year contract with Carnegie Mellon.

The story doesn't end there, though. The techniques developed during this seemingly failed project—particularly hidden Markov models, a mathematical framework for recognizing patterns in sequential data—became foundational to commercial speech recognition. By 2001, the speech recognition market had grown to four billion dollars. Technology dismissed as inadequate in the 1970s eventually powered products used by millions.

Was the First Winter Real?

Here's a provocative question: did the first AI winter, in the 1970s, actually happen?

The historian Thomas Haigh has argued that while military funding declined—largely due to congressional legislation separating military and academic activities—professional interest in AI actually grew throughout the decade. He points to membership data from ACM SIGART, the Special Interest Group on Artificial Intelligence, as evidence.

When the Lighthill report was published in 1973, SIGART had 1,241 members—already double the 1969 figure. The next five years are supposedly the darkest period of the first AI winter. But by mid-1978, SIGART membership had nearly tripled to 3,500. The group was growing faster than ever, outpacing even the Association for Computing Machinery as a whole. One in every eleven ACM members belonged to SIGART.

This suggests something more nuanced than a simple winter. Military funding collapsed, yes. But academic and commercial interest kept growing. Perhaps "winter" is too simple a metaphor for what was really a complex redistribution of attention and resources.

The Expert System Boom and Bust

The 1980s brought a new wave of AI enthusiasm, this time centered on "expert systems"—programs that encoded human expertise as rules a computer could follow. The idea was to capture the knowledge of specialists in medicine, engineering, finance, or other fields and make that expertise widely available.

The first commercial success story was XCON, developed at Carnegie Mellon for Digital Equipment Corporation. XCON configured computer systems, a task that previously required skilled engineers. It worked remarkably well. Over six years, the system was estimated to have saved the company forty million dollars.

Corporations worldwide rushed to develop their own expert systems. By 1985, businesses were spending over a billion dollars annually on AI, primarily on in-house AI departments building expert systems. An entire industry emerged to support this boom. Software companies like Teknowledge and Intellicorp sold development tools. Hardware companies like Symbolics and LISP Machines Inc. built specialized computers optimized to run LISP, the programming language favored by American AI researchers.

Then, in 1987, the market collapsed.

Desktop workstations from companies like Sun Microsystems had grown powerful enough to run AI software without specialized hardware. Why buy an expensive LISP machine when a general-purpose workstation could do the job? The specialized AI hardware market imploded almost overnight. Companies that had seemed like the future of computing suddenly faced extinction.

The expert systems themselves also ran into trouble. Maintaining them was unexpectedly difficult. The rules that encoded human expertise needed constant updating as circumstances changed. Many systems that worked well initially became obsolete as the domains they modeled evolved. Throughout the 1990s, corporations quietly abandoned expert systems they had spent fortunes building.

What Ends an AI Winter?

Winters end when something genuinely works.

Neural networks revived when backpropagation made them trainable. Speech recognition eventually produced commercial products. Even the abandoned expert systems proved that AI could create genuine business value—the problem was sustainability, not fundamental capability.

The current AI boom, which began around 2012 and accelerated dramatically through the 2020s, follows the same pattern. Deep learning—neural networks with many layers, trained on massive datasets using powerful computers—started winning competitions and solving problems that had resisted decades of effort. Image recognition, language translation, game playing, and eventually text generation reached levels that seemed impossible just years earlier.

This success was real, not hype. That's what distinguishes a sustainable advance from a bubble destined to pop.

The Lessons of the Winters

What can we learn from AI's boom-and-bust cycles?

First, researchers consistently underestimate how hard intelligence really is. Problems that seem straightforward—understanding language, recognizing objects, navigating the world—turn out to require vast amounts of implicit knowledge that humans absorb without conscious effort. The "commonsense knowledge problem" that defeated early machine translation remains relevant today, even as language models grow vastly more capable.

Second, the gap between laboratory demonstrations and real-world applications is enormous. Systems that work on carefully constructed test cases often fail when confronted with the messy complexity of actual use. The combinatorial explosion Lighthill identified in 1973 still lurks behind many AI limitations.

Third, hype creates backlash. Overpromising isn't just intellectually dishonest—it's strategically destructive. The researchers who inflated expectations in the 1960s and 1970s didn't just disappoint their funders; they damaged the field for everyone who came after.

Fourth, useful technology survives. The speech recognition work DARPA cancelled in the 1970s eventually powered a four-billion-dollar industry. Neural networks abandoned for over a decade became the foundation of modern AI. Expert systems fell out of fashion, but the basic idea of encoding knowledge as rules persists in various forms. What works endures, even through winters.

Are We in a Bubble?

The question hovering over today's AI boom is whether another winter is coming. Billions of dollars flow into AI companies. Expectations run extraordinarily high. The pattern that preceded every previous bust is clearly visible.

But there are differences. Current AI systems produce genuine value at scale. Hundreds of millions of people use AI-powered tools daily. The gap between hype and reality, while still present, is narrower than in previous cycles.

Previous winters didn't stop progress—they just slowed funding and concentrated research among true believers who continued working regardless. If another winter comes, it will likely follow the same pattern. Some companies will fail. Some research directions will be abandoned. But the underlying capabilities won't disappear, and eventually—as always—something will work well enough to thaw the frost.

Roger Schank and Marvin Minsky were right in 1984. The winter came. But they were also, in a deeper sense, wrong. AI didn't end. It just went through a cycle as old as the field itself, and emerged stronger on the other side.

That's the real lesson of the AI winters. Not that enthusiasm is dangerous—though it can be—but that the field endures. Each winter kills the weakest ideas and starves the fraudulent claims. What survives is genuine. And when spring comes, it builds on everything the winter couldn't destroy.