Source: https://www.wikiwand.com/en/AI_winter (via https://en.wikipedia.org/wiki/AI_winter)
An AI winter represents a period characterized by diminished funding and reduced research interest in artificial intelligence. The field has experienced multiple cycles of excessive optimism followed by disillusionment, funding reductions, and eventual revival after significant time gaps.
The term emerged publicly in 1984 during an AAAI annual meeting debate. Leading researchers Roger Schank and Marvin Minsky warned that 1980s enthusiasm had become unsustainable and predicted a "nuclear winter"-like cascade: community pessimism → media negativity → severe funding cuts → research stagnation.
Machine Translation (1966)
Natural language processing originated in the 1930s, gaining momentum after Warren Weaver's 1949 memorandum on machine translation. The Georgetown-IBM demonstration in 1954 generated excitement but involved only "49 Russian sentences with a 250-word vocabulary." Media coverage proved exaggerated relative to actual capabilities.
The field underestimated word-sense disambiguation challenges. The Automatic Language Processing Advisory Committee (ALPAC) concluded in 1966 that machine translation was "more expensive, less accurate and slower than human translation" despite $20 million investment. Support ended, careers were destroyed.
Neural Networks (1969)
Frank Rosenblatt's perceptrons initially showed promise but faced criticism from Marvin Minsky and Seymour Papert's influential 1969 book highlighting computational limitations. While multilayered perceptrons existed theoretically, no one knew how to train them—backpropagation remained undiscovered. Major funding dried up throughout the 1970s-early 1980s until John Hopfield and David Rumelhart revived the field mid-1980s.
The Lighthill Report
British professor Sir James Lighthill's 1973 evaluation criticized AI's failure to achieve objectives and identified the "combinatorial explosion" problem—algorithms that worked on toy problems failed catastrophically on real-world scale. This led to complete dismantling of UK AI research, with only Edinburgh, Essex, and Sussex continuing work. Large-scale revival didn't occur until 1983 when the government-backed Alvey project allocated £350 million.
DARPA Funding Cuts
The 1969 Mansfield Amendment redirected DARPA toward "mission-oriented direct research" rather than basic undirected work. AI researchers faced impossibly high standards. Hans Moravec blamed colleagues' "increasing exaggeration" of initial promises, noting staff felt compelled to promise more with each proposal despite consistently underdelivering. Contracts worth millions faced near-elimination as punishment.
One exception: the Dynamic Analysis and Replanning Tool proved enormously successful, saving billions during the first Gulf War and justifying DARPA's pragmatic approach.
Speech Understanding Research Debacle
DARPA expected a system responding to pilot voice commands. The Carnegie Mellon team delivered recognition requiring "words spoken in a particular order." Feeling deceived, DARPA cancelled a three-million-dollar annual contract in 1974. Ironically, the underlying technology (hidden Markov models) later proved valuable, with the speech recognition market reaching $4 billion by 2001.
Contrary Perspective
Historian Thomas Haigh argues the "AI winter" narrative oversimplifies 1970s activity. ACM's SIGART membership nearly tripled from 1973-1978 (1,241 to 3,500 members), growing faster than ACM overall. Professional interest demonstrably increased despite redirected military funding.
LISP Machine Market Collapse
Expert systems like XCON saved Digital Equipment Corporation an estimated $40 million over six years, sparking corporate adoption. By 1985, companies spent over $1 billion on AI, primarily internal departments. Specialized LISP machine manufacturers like Symbolics and LISP Machines Inc. flourished.
By 1987, Sun Microsystems workstations and portable LISP implementations offered superior alternatives. "An entire industry worth half a billion dollars was replaced in a single year." Early 1990s saw most commercial LISP companies fail, including Symbolics and Lucid Inc.
Expert Systems Decline
Early successful systems proved too expensive to maintain. They couldn't learn, made "grotesque mistakes" with unusual inputs, and suffered from identified problems like the qualification problem. Expert systems remained useful only in narrow contexts, forcing remaining shell companies to downsize and explore new paradigms like case-based reasoning.
Fifth Generation Project Failure
Japan's 1981 Ministry initiative allocated $850 million targeting conversation capability, language translation, image interpretation, and human-like reasoning. By June 1992, these ambitious 1981 goals remained unmet, ending "not with a successful roar, but with a whimper."
Strategic Computing Initiative Cutbacks
DARPA resumed AI funding in 1983 through the Strategic Computing Initiative, reaching $100 million by 1985 across 92 projects at 60 institutions. Leadership change proved catastrophic: Jack Schwarz, ascending to IPTO leadership in 1987, dismissed expert systems as "clever programming" and cut funding "deeply and brutally," claiming DARPA should "surf" rather than "dog paddle" through emerging technologies. Few projects survived; DART battle management succeeded, but autonomous vehicles and pilot assistants never materialized.
AI terminology itself became radioactive. Researchers deliberately adopted alternative labels—"informatics," "machine learning," "analytics," "knowledge-based systems," "business rules management," "cognitive systems," "intelligent agents," "computational intelligence"—to secure funding by "avoiding the stigma of false promises attached to the name 'artificial intelligence.'"
Contemporary industry commentary reflected persistent skepticism. One 2007 Economist quote noted investors avoided "voice recognition" due to associations with systems that "failed to live up to their promises." A 2005 New York Times article observed some computer scientists "avoided the term artificial intelligence for fear of being viewed as wild-eyed dreamers."
Yet AI technology quietly became embedded in larger systems. Nick Bostrom explained in 2006 that "cutting edge AI has filtered into general applications, often without being called AI because once something becomes useful enough and common enough it's not labeled AI anymore." Rodney Brooks similarly noted "there's this stupid myth out there that AI has failed, but AI is around you every second of the day."
AI has reached historically unprecedented interest and investment levels across publications, patent applications, and job openings. Investment reached $50 billion in 2022, with predictions of $364 billion by 2025. Approximately 800,000 U.S. job openings existed in 2022.
Breakthrough successes include Google Translate, ImageNet-powered image recognition, game-playing systems (AlphaZero, AlphaGo, Watson), and deep learning networks. The critical 2012 turning point occurred when AlexNet won the ImageNet Large Scale Visual Recognition Challenge with error rates half those of second-place competitors.
OpenAI's ChatGPT release in late 2022 reignited widespread AI discussion. The chatbot surpassed 100 million users by January 2023, reinvigorating discourse about artificial intelligence's societal implications.