If you only know artificial intelligence as a technology on an unstoppable upward trajectory, the phrase “AI winter” can come as a surprise. Yet for roughly two decades of its history, the field of AI experienced repeated collapses in funding, public interest, and academic credibility so severe that many researchers abandoned the term altogether. Understanding the AI winters — plural, because there were two major ones — is essential to understanding why the field’s progress has never been as linear as the headlines suggest.
Defining the AI Winter
An “AI winter” refers to a period of reduced funding, diminished interest, and widespread disillusionment with artificial intelligence research, typically following a period of inflated expectations that the technology failed to meet. The term is a deliberate echo of “nuclear winter,” borrowed to convey just how damaging and prolonged these downturns felt to the researchers living through them. Winters are not simply slow periods — they are sharp corrections that follow a boom, driven by a widening gap between what was promised and what was delivered.
The Setup: Early Overconfidence
To understand why the winters happened, it helps to look at what came before them. In the 1950s and 1960s, pioneers of AI made strikingly confident predictions. Some researchers suggested that machines capable of doing any work a human could do would exist within a generation, and that fundamental problems of machine intelligence would be largely solved within a decade. These predictions were not idle talk — they were used to justify substantial government funding, particularly from defense agencies in the United States and the United Kingdom, which hoped AI could deliver breakthroughs in machine translation, automated reasoning, and robotics.
The First AI Winter (Mid-1970s)
By the early-to-mid 1970s, the gap between promise and reality had become impossible to ignore. Machine translation programs, an early showcase for AI funding, had produced embarrassingly bad results — famously, a program was said to have translated “the spirit is willing but the flesh is weak” into Russian and back into English as something resembling “the vodka is good but the meat is rotten,” illustrating how far the technology was from genuine language understanding. More broadly, researchers discovered that problems which seemed simple for computers (like recognizing objects in images or understanding ordinary sentences) were in fact enormously difficult, while problems that seemed hard for humans (like complex mathematical calculations) were comparatively easy for machines. This mismatch, now informally known as Moravec’s paradox, caught early researchers off guard.
A pivotal moment came in 1973 with the Lighthill Report, commissioned by the UK’s Science Research Council and authored by mathematician Sir James Lighthill. The report concluded that AI research had failed to achieve its grand, headline-grabbing objectives, and it was sharply critical of the field’s progress on general-purpose, broadly intelligent machines. In response, the British government drastically cut AI funding, and the report’s influence spread internationally. In the United States, the Defense Advanced Research Projects Agency (DARPA) grew similarly disillusioned with speech recognition and other AI projects that had failed to deliver practical results, leading to substantial funding cuts there as well. Academic departments scaled back AI programs, and many computer scientists quietly moved into other subfields to avoid being associated with an area that had lost credibility.
A Brief Recovery: Expert Systems in the 1980s
The field found new life in the early 1980s through expert systems — programs built to encode the specialized knowledge of human experts into sets of if-then rules for narrow, well-defined tasks such as medical diagnosis or chemical analysis. Unlike the grand ambitions of earlier AI research, expert systems promised modest, practical value, and companies around the world began investing heavily. By the mid-1980s, the AI industry had grown into a market worth well over a billion dollars annually, with specialized hardware companies building computers optimized to run these rule-based systems. Japan’s government-backed Fifth Generation Computer Systems project, launched in 1982, further fueled a global race to invest in AI, as other nations feared falling behind.
The Second AI Winter (Late 1980s–Early 1990s)
The expert-systems boom proved just as fragile as the first wave of AI optimism. These systems were expensive to build, difficult to update as knowledge changed, and completely unable to generalize beyond their narrow, pre-programmed domains — a system built to diagnose infections couldn’t be adapted to a new problem without being rebuilt largely from scratch. Specialized “Lisp machine” hardware companies, built specifically to run AI software, collapsed almost overnight in 1987 as cheaper, more powerful general-purpose computers from companies like Sun Microsystems and Apple made dedicated AI hardware obsolete. The Strategic Computing Initiative, a major source of US defense funding for AI, cut its support as results failed to materialize at the promised scale. Once again, investment dried up, research budgets shrank, and the term “artificial intelligence” became something of a liability on grant applications. Many researchers began describing their work using alternative labels — “machine learning,” “pattern recognition,” or “informatics” — specifically to distance themselves from AI’s damaged reputation.
What Finally Ended the Winters
The thaw that followed the second AI winter was gradual rather than dramatic, and it came from several directions at once. Researchers shifted toward narrower, statistically grounded approaches that could demonstrate measurable, verifiable progress rather than chasing broad claims of human-level intelligence. Steady improvements in computing power, following the trajectory of Moore’s Law, made more sophisticated algorithms computationally feasible. The 1997 defeat of world chess champion Garry Kasparov by IBM’s Deep Blue became a widely publicized proof point that machines could genuinely outperform humans in demanding tasks, helping to rebuild public confidence. Over the following two decades, an explosion of digital data, combined with breakthroughs in neural network design and the use of powerful graphics processing units for computation, created the conditions for the deep learning boom of the 2010s — a resurgence that has, so far, not slipped back into winter.
Could There Be a Third AI Winter?
The question of whether another AI winter could occur remains a live debate among researchers and industry observers. Critics point out that today’s generative AI systems, despite their impressive capabilities, still struggle with reliability, factual accuracy, and reasoning in ways that echo the limitations exposed in earlier eras — and that enormous sums of investment capital are once again riding on promises of transformative, near-term breakthroughs. Optimists counter that today’s AI is already delivering measurable economic value across many industries, unlike the comparatively fragile expert systems of the 1980s, making a repeat collapse less likely. Whichever view proves correct, the history of the AI winters offers a durable lesson: in artificial intelligence, hype and funding tend to move together, and when the gap between promise and delivery grows too wide, the consequences can be swift and severe.