B1 · Introduction to Artificial Intelligence

AI Winters and Revivals

An AI winter is a period when funding and interest in AI collapse after promises go unmet. There were two major ones: from about 1974 to 1980, and from about 1987 to the mid-1990s.

6 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

An AI winter is a period when funding and interest in AI collapse after promises go unmet. There were two major ones: from about 1974 to 1980, and from about 1987 to the mid-1990s.

In simple words

If you promise your investors a flying car in two years and deliver a slightly better bicycle, they stop paying. AI promised too much, twice, and the money and attention dried up both times.

Why it matters

Knowing this history helps you judge today's claims. It doesn't prove another winter is coming, but it shows why promises should be measured against results.

Key terms

AI winter
A period when money and interest in AI fall sharply because results did not match promises.
Hype
Excitement and claims that run ahead of what a technology can actually do.
Expert system
A program that stores a human expert's knowledge as rules and uses them to give advice.
Combinatorial explosion
The way the number of possibilities grows so fast that checking them all becomes impossible.
Lighthill Report
A 1973 report for the British government that criticised AI research and led to funding cuts.

Explained in detail

The first winter, about 1974 to 1980

In the 1950s and 1960s governments funded AI generously, expecting quick results. One hope was automatic translation between Russian and English. After ten years of work, a committee of American scientists reported in 1966 that machine translation was slower, less accurate and more expensive than human translators, and most of the funding stopped. In 1969 Marvin Minsky and Seymour Papert published a book showing that the simple Perceptron could not learn some very basic patterns, and research on neural networks slowed for years.

In 1973 the mathematician James Lighthill wrote a report for the British government. He concluded that AI had failed to achieve its "grandiose objectives" and that methods which worked on small toy problems broke down on real ones, because the number of possibilities to check grew impossibly fast, a difficulty known as combinatorial explosion. Britain cut most AI funding, and the United States also reduced its support. Computers of the time were simply too slow and had too little memory for what researchers wanted to do.

The revival: expert systems

In the 1980s AI returned in a more modest and practical form. Expert systems captured the knowledge of human specialists as hundreds or thousands of rules. One famous system, used by the computer maker Digital Equipment Corporation to work out which parts each customer order needed, was reported to save the company tens of millions of dollars a year. Companies rushed to build their own. In 1982 Japan announced the Fifth Generation Computer project, a ten-year national plan to build intelligent computers, and other countries increased their spending to keep up.

The second winter, about 1987 to the mid-1990s

Expert systems turned out to be brittle. They worked inside a narrow area and gave absurd answers just outside it. Every new piece of knowledge had to be added by hand, and as the rule collections grew, rules began to conflict and the systems became expensive to maintain. They could not learn. In 1987 the market for the special computers built to run AI software collapsed, as ordinary desktop computers became powerful enough and far cheaper. Japan's Fifth Generation project ended without reaching its main goals. Funding dried up again, and for some years researchers avoided the words "artificial intelligence" when asking for money.

How AI came back

The recovery was quiet. Researchers turned from hand-written rules to methods that learn from data using statistics. The internet began to supply enormous amounts of text and images to learn from, and computers kept getting faster. Useful products appeared without much fanfare: spam filters, search engines, speech recognition on telephones, recommendation systems in online shops. In 1997 IBM's Deep Blue beat the world chess champion. Then, in 2012, deep learning produced results that surprised even its supporters, and the present boom began.

The pattern to remember

  • A new method produces impressive results on a limited problem.
  • Researchers, companies and the press promise far more than it can yet do.
  • Money pours in on the strength of those promises.
  • The method meets its limits, and disappointment follows.
  • The useful parts survive quietly and become ordinary tools.

How it works

WinterMain causesWhat ended it
First, about 1974–80Machine translation disappointed, the 1973 Lighthill report in the UK was critical, and computers were too weakExpert systems began to make money
Second, about 1987–mid-1990sExpert systems proved costly and brittle, and the market for specialised AI machines collapsedStatistical machine learning and the internet's data

The lesson

Winters didn't stop the science. Many of today's key ideas, such as backpropagation and convolutional networks, were developed through these difficult years by people who kept working.

Try it yourself

The rise and fall of excitement

Watch what happens when promises run ahead of results.

Promise compared with result

Did each event help bring an AI winter, or help AI come back?

  • The internet produced huge amounts of data

  • A UK report said AI had not met its goals (1973)

  • Computers became far faster and cheaper

  • A neural network won the ImageNet contest by a wide margin (2012)

  • Expert systems proved costly and hard to keep up to date

  • Machine translation fell far short of promises (1966 report)

Common mistakes

  • Believing either "this time is exactly the same" or "this time is completely different". Look at the actual results.
  • Treating hype as progress.

Check yourself

Question 1

1. What caused the AI winters?

Question 2

2. Which report hurt UK AI funding in 1973?

Summary

  • Two major winters: the 1970s and the late 1980s.
  • Over-promising led to funding collapses.
  • Key research carried on quietly through them.

Finished reading?

Sources: Russell & Norvig, Artificial Intelligence: A Modern Approach, ch. 1