# What was the first AI winter, and why did the money for AI dry up in the 1970s?

Date: 2019-11-20 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-was-the-first-ai-winter/

**Answer:** The first AI winter was a stretch of roughly 1974 to 1980 when government funding and public enthusiasm for artificial intelligence fell sharply in the United States and Britain. It followed years of bold promises that early programs could not keep, a critical British review known as the Lighthill report, and a shift at the U.S. defense research agency DARPA toward projects with clear military goals.

In the 1960s, artificial intelligence felt unstoppable. Programs proved theorems, played checkers and held simple conversations, and leading researchers predicted thinking machines within a generation. Then, in the first half of the 1970s, the mood turned. Funders lost patience, budgets were cut, and by 1974 money for AI projects had become hard to find. The slump lasted until about 1980.

People now call such a period an **AI winter**: a time of reduced funding and interest in AI research, by analogy with the idea of a "nuclear winter." The name itself is younger than the slump. It was first used in 1984, when a public debate on the subject was held at the yearly conference of the American Association of Artificial Intelligence (AAAI). It was coined by researchers who had lived through the 1970s cuts and feared a repeat.

## When and where it happened

There was no single date. The first winter was a set of decisions on both sides of the Atlantic, spread across a few years:

| Year | What happened | Effect |
|---|---|---|
| 1966 | The ALPAC report in the U.S. finds machine translation slower, less accurate and costlier than human translators | U.S. government support for translation research ends |
| 1969 | Minsky and Papert's book *Perceptrons* shows the limits of simple neural networks | Neural network research nearly stops for about a decade |
| 1969 | The Mansfield Amendment requires DARPA to fund mission-oriented research | Open-ended "blue sky" AI grants dry up |
| 1973 | The Lighthill report in Britain judges that AI has failed to deliver | British AI funding is cut back to a few universities |
| 1974 | DARPA cancels a three-million-dollar-a-year grant to Carnegie Mellon's speech project | By 1974, money for AI is hard to find |

In the United States, most AI money came from DARPA, the Defense Department's research agency (then called ARPA). Through the 1960s it had given millions of dollars to the leading AI labs with few strings attached, trusting their leaders to choose good problems. After the Mansfield Amendment, it wanted projects with identifiable military goals. It was also unhappy with Carnegie Mellon's Speech Understanding Research program: DARPA had hoped for a system that could take spoken commands from a pilot, and got one that understood English only when the words came in a fixed order.

In Britain, the Science Research Council commissioned the mathematician Sir James Lighthill to review the field. His 1973 report concluded that in no part of AI had discoveries produced the major impact that had been promised. Afterward, serious AI research continued in only a handful of places, including Edinburgh, Essex and Sussex.

## Why the early promises failed

The disappointment had real technical causes, and researchers of the 1970s named most of them themselves.

**Computers were far too weak.** Memory and speed were tiny by later standards. One celebrated language program, by Ross Quillian, was shown off with just twenty words of vocabulary, since no more would fit in memory. In 1976 the roboticist Hans Moravec estimated that computers would need to become millions of times more powerful before they could show intelligence; the fastest supercomputer of that year, the Cray-1, managed around 80 to 130 million instructions per second.

**The combinatorial explosion.** Lighthill's main charge was that the number of possibilities a program must consider grows explosively as a problem gets bigger. Methods that worked on small puzzles became hopeless on real-world ones.

**Common sense is enormous.** Understanding a sentence or a picture needs a huge store of everyday knowledge, the kind every child has. Nobody in 1970 could build a database that large, and nobody knew how a program could learn it.

**Easy for people, hard for machines.** Proving a theorem turned out to be easier for a computer than recognizing a face or walking across a room without bumping into things. This puzzle, later called Moravec's paradox, explains why vision and robotics made so little progress.

**Promises outran results.** Moravec blamed his colleagues for a cycle of exaggeration: having promised too much in one proposal, they felt they could not promise less in the next.

> **Try it:** Find a prediction about AI made in the 1960s or 1970s (Lighthill's report quotes several, such as machines exceeding human intelligence by 2000). Write down what was promised, by when, and what the gap turned out to be. Then do the same for one confident AI prediction you have read this year.

## What survived the winter

Research did not stop; it changed. Funding moved to work with practical, narrow goals, and that shift pushed researchers toward knowledge-heavy programs. Lighthill had noted that the successful programs, such as Stanford's DENDRAL chemistry system, relied on deep knowledge of one domain. That insight drove the expert systems of the late 1970s, including Stanford's MYCIN for infections. New ideas also grew in logic programming, through the language Prolog created in Marseille, and in research on commonsense reasoning.

By 1980 the winter was ending. Digital Equipment Corporation put a rule-based system called XCON to work configuring computer orders, and businesses began investing in expert systems. That boom later brought its own bust: a second winter in the late 1980s and early 1990s, when expert systems proved costly to maintain and the market for specialized Lisp machines collapsed.

## Could another AI winter happen?

It could, and the pattern is worth knowing. One account describes AI winters as a chain reaction: pessimism among researchers spreads to the press, funders cut back sharply, and serious research shrinks. The trigger in each past case was a gap between what had been promised and what was delivered.

In late 2019 the situation differs in important ways. Much of today's funding comes from large technology companies that use AI in products people already rely on every day, such as search, translation, photo tagging and speech recognition, rather than from a single government agency. Computers are vastly faster than the machines of 1974, and neural networks, written off after *Perceptrons*, now drive most of the field's progress.

Still, some commentators warn of overpromising around self-driving cars and claims of human-level AI. A sensible stance for a beginner is to watch the gap between demonstrations and dependable, everyday use. When that gap narrows, confidence is earned; when it widens while promises keep growing, the conditions for a cold spell return.

## Sources

- Artificial Intelligence: A General Survey: http://www.chilton-computing.org.uk/inf/literature/reports/lighthill_report/p001.htm
- AI winter: https://web.archive.org/web/20190501000000/https://en.wikipedia.org/wiki/AI_winter
- History of artificial intelligence: https://web.archive.org/web/20190301000000/https://en.wikipedia.org/wiki/History_of_artificial_intelligence