AI winter

• 1. Introduction

  ›

  AI winter is the name researchers gave to something that kept happening: a surge of excitement about artificial intelligence, followed by crushing disappointment, followed by funding cuts so severe the field nearly went dark. The term itself first appeared in 1984, at the annual meeting of the American Association of Artificial Intelligence. Two of the field's most prominent figures, Roger Schank and Marvin Minsky, stood up and warned their colleagues that history was about to repeat itself. They had lived through the slowdown of the 1970s and saw the same dangerous optimism building again. Their warning was vivid: they described a chain reaction modeled on the concept of nuclear winter, where pessimism would spread from researchers to the press, then trigger funding cuts so severe that serious research would effectively stop. Three years after they gave that warning, the billion-dollar AI industry began to collapse. The field would experience two major winters, roughly 1974-1980 and 1987-2000, plus a string of smaller crises stretching back to 1966. What caused these collapses? Why did they keep happening? And what finally broke the cycle?

• 2. The Machine That Could Not Translate

  ›

  In 1954, a joint demonstration by Georgetown University and IBM captured headlines across the world. Newspapers described the machine as a bilingual brain, a polyglot brainchild, a robot that could translate Russian into English. The actual demonstration translated exactly 49 Russian sentences using a vocabulary of just 250 words. For context, a 2006 study by Paul Nation found that humans need roughly 8,000 to 9,000 word families just to understand written text with 98% accuracy. The Georgetown-IBM machine was operating at a fraction of that. The gap between the headlines and the reality of that curated demonstration captures everything about what would go wrong.

  The US government, driven by Cold War anxieties, had been pouring money into machine translation since 1954. The CIA believed strongly in the technology and funded it partly for intelligence purposes, and partly because they recognized that its implications reached far beyond spy work. Noam Chomsky's new work on grammar seemed to promise a streamlined path forward. Researchers were making confident predictions about imminent breakthroughs. What they had failed to reckon with was something they would later call the commonsense knowledge problem. A machine that does not understand the meaning of a sentence cannot translate it correctly. A Russian phrase meaning "the spirit is willing but the flesh is weak" reportedly came back, after a round trip through a translation system, as "the vodka is good but the meat is rotten."

  By 1964, the National Research Council had grown impatient and formed the Automatic Language Processing Advisory Committee to investigate. Their 1966 report concluded flatly that machine translation was more expensive, less accurate, and slower than simply hiring human translators. The NRC ended all support after spending roughly 20 million dollars. Careers ended. Research stopped. The same path from enthusiasm to collapse would be traced again and again across different corners of the field.

• 3. Perceptrons and the Limits of Promise

  ›

  Frank Rosenblatt invented the perceptron and defended it with something his contemporaries described as the sheer force of his personality. He predicted publicly that the perceptron might eventually learn, make decisions, and even translate languages. The enthusiasm was real, but so was the problem: nobody in the 1960s knew how to train a network with more than one layer. The mathematical technique called backpropagation, which would eventually make that possible, was still years away.

  In 1969, Marvin Minsky and Seymour Papert published a book called Perceptrons. It laid out with mathematical precision what single-layer networks could not do. The book did acknowledge that multilayered perceptrons were not subject to the same limitations, but that acknowledgment arrived alongside the uncomfortable fact that no one knew how to train them. The effect on funding was immediate. Major support for neural network research became difficult to find throughout the 1970s and into the early 1980s. Important theoretical work continued despite the drought, and by the mid-1980s researchers like John Hopfield and David Rumelhart had revived large-scale interest in the approach.

  Rosenblatt did not live to see the vindication. He died in a boating accident shortly after Perceptrons was published.

• 4. The Lighthill Report and DARPA's Reversal

  ›

  In 1973, the UK Parliament asked professor Sir James Lighthill to evaluate what AI research in Britain had actually accomplished. His verdict was withering. The field had failed to achieve its grandiose objectives. Nothing it was doing, Lighthill argued, could not be done better under the banner of some other science. He singled out what he called the problem of combinatorial explosion: many of AI's most celebrated algorithms worked only on scaled-down, toy versions of real problems. Put them in front of the real world, and they would grind to a halt.

  The report was challenged in a debate broadcast on the BBC. Lighthill faced Donald Michie, John McCarthy, and Richard Gregory in a program called "The general purpose robot is a mirage." McCarthy later wrote that the combinatorial explosion problem had been recognized within AI from the very beginning. The debate did not save British AI funding. Research was effectively dismantled across the country, surviving only at Edinburgh, Essex, and Sussex. It would not revive at scale until 1983, when a British government project called Alvey began funding the field again from a war chest of 350 million pounds, motivated partly by competition with Japan's Fifth Generation computer project.

  Across the Atlantic, a parallel reversal was unfolding. During the 1960s, the agency then known as ARPA had funded AI research generously and with few conditions. J. C. R. Licklider, the founding director of DARPA's computing division, believed in funding people rather than projects. That philosophy changed after the Mansfield Amendment passed in 1969, which required the agency to fund only mission-oriented research with practical military applications. AI proposals now had to demonstrate near-term usefulness. By 1974, that standard was nearly impossible to meet. Hans Moravec, one of the researchers who lived through it, described a trap that many of his colleagues had fallen into: early promises to DARPA had been too optimistic, the results had fallen short, but researchers felt they could not promise less in the next proposal, so they promised more. Eventually, some staff at DARPA had simply lost patience. Moravec quoted an official as saying that some of these people were going to be taught a lesson by having their two-million-dollar-a-year contracts cut to almost nothing.

• 5. Expert Systems Rise and Fall

  ›

  XCON was the first commercial expert system, built at Carnegie Mellon for Digital Equipment Corporation. It was, by any measure, a genuine success: estimates put its savings for the company at 40 million dollars over just six years. By 1985, corporations worldwide were spending over a billion dollars a year on AI, most of it flowing into internal AI departments building their own expert systems. A whole ecosystem had grown up to support them: software companies like Teknowledge and Intellicorp, and hardware companies like Symbolics and LISP Machines Inc., which built specialized computers optimized to run the programming language LISP.

  Minsky and Schank had warned in 1984 that the boom would collapse. In 1987, three years after that prediction, it did. General-purpose workstations from companies like Sun Microsystems had become powerful enough to run LISP efficiently, and desktop computers from Apple and IBM had reached similar capability by the same year. There was simply no longer a reason to buy an expensive machine built exclusively to run one language. An entire industry valued at half a billion dollars was replaced in a single year.

  The expert systems themselves followed a few years later. By the early 1990s, systems like XCON had become too expensive to maintain. They were brittle, breaking down unpredictably when confronted with unusual inputs. They could not learn from new information. They fell into problems that researchers in nonmonotonic logic had identified years earlier. The Japanese Ministry of International Trade and Industry had set aside 850 million dollars in 1981 for the Fifth Generation computer project, with goals including machines that could hold conversations, translate languages, and reason like humans. By the 1st of June 1992, according to HP Newquist in The Brain Makers, the project ended not with a successful roar, but with a whimper.

• 6. A Stigma That Outlasted the Crisis

  ›

  By the early 2000s, AI as a label had become something researchers actively avoided. A 2005 piece in the New York Times quoted John Markoff observing that some computer scientists avoided the term "artificial intelligence" for fear of being seen as wild-eyed dreamers. A 2006 piece in Pittsburgh Business Times quoted Patty Tascarella noting that the word "robotics" carried a stigma that could hurt a company's chances at funding. Alex Castro, quoted in The Economist in 2007, described investors being put off by the phrase "voice recognition" precisely because, like "artificial intelligence," it was associated with repeated failures to deliver.

  Researchers responded by renaming their work. Informatics, machine learning, analytics, knowledge-based systems, cognitive systems, intelligent agents: all of these terms flourished in the mid-2000s partly because they described specific tools or specific sub-problems, and partly because they carried none of the baggage attached to AI. Nick Bostrom explained in 2006 that a great deal of cutting-edge AI had quietly filtered into general applications without being called AI, because once something becomes useful enough and common enough, it stops being labeled AI at all. Rodney Brooks, around the same time, put it more bluntly: there was a stupid myth, he said, that AI had failed, when in fact AI was around people every second of the day.

  Historian Thomas Haigh has made a broader challenge to the standard narrative. Using membership counts in ACM's SIGART as a proxy for research activity, he found that when the Lighthill report was published in 1973, the group had 1,241 members, roughly double the 1969 level. By mid-1978, membership had nearly tripled again to 3,500. By that measure, one in every eleven ACM members was in SIGART, and the rate of growth was accelerating, not contracting.

• 7. The Current AI Spring

  ›

  A turning point arrived in 2012, when a deep learning network called AlexNet won the ImageNet Large Scale Visual Recognition Challenge with half as many errors as the second-place finisher. That single result helped shift the direction of an entire field. Investment and interest in artificial intelligence climbed steadily through the 2010s, and in the 2020s reached levels unmatched at any earlier point in the field's history. Total investment stood at 50 billion dollars in 2022. Large tech companies were projected to spend 364 billion dollars in 2025. U.S. job openings related to AI reached 800,000 in 2022.

  The successes driving that investment were real and measurable: advances in language translation through services like Google Translate, image recognition commercialized through Google Image Search, and game-playing systems that had defeated world champions. AlphaGo beat the world's best Go players. AlphaZero became a chess champion. Watson won Jeopardy. In late 2022, OpenAI released ChatGPT. By January 2023, the chatbot had passed 100 million users. The speech recognition technology that DARPA cancelled in 1974, after cutting a three-million-dollar-a-year contract at Carnegie Mellon, had quietly continued developing. The market for commercial speech recognition systems built on techniques from that era reached 4 billion dollars by 2001, long before the current boom. Whether the current spring will hold, or whether the pattern that Minsky and Schank described in 1984 still applies, remains an open question that the history documented here is uniquely positioned to inform.