Concept · Chapter 1: What Is Artificial Intelligence?
AI Winters
AI winters were periods when inflated expectations collided with limited results, and funding and interest in AI collapsed for years.
The problem
Early AI researchers made confident predictions that the technology of the time could not deliver.
The solution
There was no quick fix: funders, governments and companies pulled back, and interest returned only when new approaches produced measurable results.
The consequence
The field became wary of overpromising, and work shifted toward narrower, measurable problems and statistical methods — the path that eventually led to modern machine learning.
Two winters
The first (≈1974–1980). Early programs that worked on toy problems didn't scale to real ones: machine translation disappointed (the 1966 ALPAC report), perceptrons were shown to be limited (1969), and in 1973 the UK's Lighthill report criticized AI's failure to meet its grand objectives. Funding in the UK and US dried up.
The second (from 1987). The expert-systems boom of the early 1980s ended when specialized LISP machines were overtaken by cheaper general-purpose computers and maintaining large rule bases proved expensive. Japan's ambitious Fifth Generation project (1982–1994) missed its goals. "AI" became an unfashionable label for years, even as statistical machine learning quietly advanced.
The lesson for reading today's AI
Each winter followed a gap between what demos suggested and what systems reliably did Interpretation. That doesn't mean today's progress is hype — current systems are far more capable and widely used than anything in those eras — but it is a reason to read claims with the evidence labels this site uses: what is established, what is interpretation, and what is speculative.
What to remember
- First winter ≈ 1974–1980: after critiques like the 1973 Lighthill report.
- Second winter from 1987: expert systems and LISP-machine hardware collapse.
- Cause each time: promises far ahead of capabilities.
- Useful habit: separate demonstrated evidence from extrapolation.