The AI winters: when the hype collapsed and the money ran out
Twice, in the 1970s and late 1980s, artificial intelligence failed to live up to its promises and funding collapsed. What happened, why, and what the AI winters can teach us about today's boom.
First published · AI Digested
Today's AI boom can make it seem as if progress only goes one way. It hasn't always. Twice in its history, artificial intelligence promised the world, failed to deliver, and saw its funding and reputation collapse. These periods are known as the AI winters.
The first boom
After the 1956 Dartmouth workshop, AI researchers made impressive early progress and even more impressive predictions. Governments, especially the US military research agency now called DARPA, funded the field generously, often with few strings attached.
Researchers promised machine translation, general problem-solvers and human-level intelligence within a generation.
Cracks appear
Machine translation stalls. During the Cold War, the US government badly wanted computers to translate Russian. In 1966, a government committee, the ALPAC report, concluded that machine translation was slower, less accurate and more expensive than human translation. Funding was slashed.
Neural networks fall out of favour. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, showing that the simple single-layer neural networks of the time couldn't learn some basic patterns. Interest in neural networks largely dried up for over a decade.
The Lighthill report. In 1973, the UK's Science Research Council asked the mathematician Sir James Lighthill to assess AI research. His report was damning: AI had failed to achieve its "grandiose objectives". Its methods worked on toy problems but suffered from a "combinatorial explosion" when applied to the real world. Funding for AI in British universities was cut sharply.
In the US, DARPA shifted towards projects with clear, short-term military uses. By the mid-1970s, the first AI winter had arrived.
The second boom: expert systems
AI came back in the 1980s with a more practical pitch. Expert systems captured the knowledge of human specialists as large collections of if-then rules.
The most famous success was XCON, built for Digital Equipment Corporation, which configured customers' computer orders and was reported to save the company millions of dollars a year. Soon, companies around the world were building expert systems for everything from medical diagnosis to oil exploration.
Money poured in again. In 1982, Japan launched its ambitious Fifth Generation Computer project. The US and UK responded with programmes of their own; Britain's was called Alvey. Start-ups sold specialised computers, called Lisp machines, designed to run AI software.
The second winter
From about 1987, it all unravelled:
- Hardware: ordinary desktop computers from Apple and IBM became powerful enough to run AI software. The market for expensive Lisp machines collapsed almost overnight.
- Brittleness: expert systems couldn't handle situations their rules didn't cover, and they made strange mistakes when they met something unusual.
- Maintenance: keeping thousands of rules up to date was expensive. Experts' knowledge was harder to write down than anyone had expected.
- Unmet goals: Japan's Fifth Generation project ended in the early 1990s without achieving its headline ambitions.
"AI" became a word to avoid. Researchers rebranded their work as "machine learning", "informatics" or "knowledge-based systems" to win funding.
Quiet progress in the cold
Winters weren't dead periods. Important work continued away from the spotlight:
- Backpropagation, the method for training multi-layer neural networks, was popularised in 1986.
- Yann LeCun's neural networks were reading handwritten postcodes and cheques by the early 1990s.
- Statistical methods, learning from data rather than hand-written rules, steadily improved speech recognition and translation.
These quiet advances became the foundations of the deep learning revolution after 2012.
What the winters teach us
Hype is dangerous. Over-promising led directly to the backlash.
Toy problems are misleading. Success on neat demonstrations often failed to survive the messiness of the real world.
Timing predictions are unreliable. Pioneers expected human-level AI by the 1980s. Sceptics in the 1990s would have been astonished by today's chatbots.
The "AI effect". Each time AI solved something, like chess or route-finding, people stopped calling it AI. As the computer scientist Larry Tesler put it: "AI is whatever hasn't been done yet."
Today's boom is different in an important way: AI is in products used by hundreds of millions of people, and companies earn real revenue from it. But the scale of investment is also unprecedented, and some economists warn of a bubble. The winters are a reminder to judge AI by what it actually does, not by what's promised.
Read on
Common questions
Why is it called an 'AI winter'?
The phrase was used at a 1984 meeting of the American Association for Artificial Intelligence, where researchers warned that hype would lead to a chain reaction of pessimism, funding cuts and collapse, like a 'nuclear winter'.
Could there be another AI winter?
Some people think a slowdown is possible if AI doesn't deliver enough value to justify the enormous investment. Unlike in past winters, AI is now widely used in products that make money, which makes a full collapse less likely.
What were expert systems?
Programs that captured the knowledge of human experts as thousands of if-then rules, for tasks like diagnosing infections or configuring computer orders. They worked well in narrow areas but were expensive to build and hard to update.
Sources and further reading
Spotted a mistake or something out of date? Tell us and we will fix it. Read how we research and write our articles.
Read next
Alan Turing and the Turing test: the man who asked 'Can machines think?'
The story of Alan Turing, the British mathematician and codebreaker who imagined thinking machines in 1950, invented the Turing test, and laid the foundations of computing and AI.
Updated 22 November 2026
Dartmouth 1956: the summer artificial intelligence got its name
How a small 1956 summer workshop at Dartmouth College named the field of artificial intelligence, who was there, what they hoped to achieve, and why their optimism was so wrong about timing.
Updated 22 November 2026
Deep Blue beats Kasparov: the day a computer became chess champion
In May 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov. The story of the match, how Deep Blue worked, the controversy, and what it did and didn't prove about AI.
Updated 22 November 2026