The AI winters

When the dreams collapse

The AI winters are periods in the history of artificial intelligence when funding was drastically reduced, laboratories were closed, and the field fell into a reputational crisis. There were two such winters: the first from roughly 1974 to 1980, the second from 1987 to 1993. Both were caused by a combination of inflated expectations, technological limitations, and a gradual realization that the goals of the AI pioneers were much harder to achieve than anticipated.

The first AI winter (1974–1980)

The first winter began after the damaging Lighthill Report, published in 1973 at the request of the British government. Mathematician James Lighthill concluded that AI research was only useful for robotics and language processing, but that the ambitious claims about general intelligence and machine learning had not been realized. The report led to drastic cuts by the British government and had international repercussions. In America, DARPA and other funders heavily reduced their AI budgets. Universities disbanded AI groups; PhD students left the field.

At the same time, the limitations of the two dominant approaches became apparent. Symbolic AI — programming logical rules to simulate intelligent behavior — failed to scale to more complex problems. Neural networks — already under pressure after Minsky and Papert's attack in 1969 — could not deliver on the promises of pattern recognition.

The temporary recovery: expert systems (1980–1987)

In the early 1980s, AI seemed to return through the rise of expert systems. These were programs that encoded the knowledge of a human expert in a specific domain as a set of if-then rules and an inference engine that applied those rules. Well-known examples were MYCIN (diagnosis of bacterial infections, Stanford 1972), XCON (configuration of VAX computers, Digital Equipment Corporation), and PROSPECTOR (mineral exploration). XCON reportedly saved DEC an estimated 40 million dollars per year. The AI industry grew to a market of more than one billion dollars in 1986.

The second AI winter (1987–1993)

Expert systems, however, proved to have inherent weaknesses. They were extremely expensive to maintain: every change in domain knowledge required manual adjustment of hundreds of rules by costly knowledge engineers. They could not learn from new cases. They were brittle outside their defined domain. And they scaled poorly. When in 1987 the market for AI workstations collapsed (too expensive, too little advantage over regular computers), followed by cuts at DARPA, the second winter began.

In 1991, DARPA's Strategic Computing Initiative assessed its AI projects as disappointing and withdrew funding. Japanese plans for a "fifth generation" computer based on AI were abandoned. The word "AI" had become so tainted that companies preferred to label their products "machine learning" or "data mining."

Lessons from the winters

The AI winters taught the community valuable lessons. They showed that hype is dangerous and that excessive promises backfire. They also made clear that AI is not one technology but a collection of approaches, each with its own strengths and weaknesses. The winters also cleared the way for more solid approaches: machine learning, statistics, and — ultimately — deep learning. Ironically, the research that survived the winters laid the foundation for the breakthroughs of the 2010s.

Sources

• Lighthill, J. (1973). Artificial Intelligence: A General Survey. Science Research Council.
• Crevier, D. (1993). AI: The Tumultuous History of the Search for Artificial Intelligence. Basic Books.
• Russell, S. & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
• Wikipedia — AI winter