The First AI Winter
In the 1970s, the money and goodwill that had poured into AI dried up. You will see the reports, the mathematics and the hard problems that caused the chill, judge whether it was deserved, and find out which ideas quietly survived it.
In 1973 the BBC televised a debate at the Royal Institution in London. On one side sat Sir James Lighthill, one of Britain’s most eminent applied mathematicians, who had just written a government review of artificial intelligence. Facing him were three of the field’s leaders: Donald Michie of Edinburgh, the psychologist Richard Gregory and John McCarthy, flown in from Stanford. The researchers argued well. It made little difference. Lighthill’s report was already persuading Britain’s Science Research Council to cut most of its support for AI in universities, and similar doubts were spreading through the American agencies that funded the field. The first AI winter had begun.
Broken promises, first in translation
The trouble started with language. In 1954 a joint Georgetown–IBM demonstration had translated more than sixty Russian sentences into English using a vocabulary of about 250 words and a handful of grammar rules. Newspapers reported that fully automatic translation was a few years away, and US agencies, keen to read Soviet science, funded it generously.
A decade later the results were disappointing. The government asked the Automatic Language Processing Advisory Committee (ALPAC), chaired by John R. Pierce of Bell Labs, to take stock. Its 1966 report found machine translation slower, less accurate and more expensive than human translation, and concluded that “there is no immediate or predictable prospect of useful machine translation.” It recommended putting money into tools that help human translators and into basic research on language. American funding for machine translation shrank sharply.
Popular myth You may have heard that a system translated “The spirit is willing, but the flesh is weak” into Russian and back as “The vodka is good, but the meat is rotten.” No documented system ever produced it. The story circulated as a joke about translation long before any computer was involved.
Perceptrons and the XOR problem
In 1969 Marvin Minsky and Seymour Papert of MIT published Perceptrons, a careful mathematical study of what Rosenblatt’s single-layer learning machines could and could not compute. The headline result is easy to see with two inputs. A single-layer perceptron decides by drawing one straight line through the space of inputs: everything on one side gets “yes”, everything on the other “no”. Compare two simple functions:
AND says yes only when both inputs are 1. Plot the four cases on a square: (0,0) no, (0,1) no, (1,0) no, (1,1) yes. One straight line can cut off the single “yes” corner from the other three. A perceptron learns this easily.
XOR (“exclusive or”) says yes when exactly one input is 1: (0,1) yes, (1,0) yes, (0,0) no, (1,1) no. The two “yes” corners sit diagonally opposite each other. Try drawing one straight line that puts both yeses on one side and both nos on the other: you cannot. No setting of the weights works, so no amount of training helps.
Add a hidden layer. One hidden unit computes “at least one input is on” (OR), another computes “both are on” (AND), and the output says yes when the first fires but not the second. Multilayer networks can compute XOR and far more. The catch in 1969 was that nobody had a widely known, practical way to train the hidden layers. That would come with backpropagation.
XOR was only the simplest example. Minsky and Papert’s deeper results showed that certain kinds of single-layer perceptron could not recognise global properties such as whether a figure is connected, or whether the number of active inputs is odd, without growing impractically large. They were also sceptical, in the book’s closing pages, that multilayer versions would prove much more useful. Rosenblatt died in a boating accident in 1971, and neural-network research lost its most energetic champion.
Debated How much Perceptrons itself killed neural networks is still argued over.
- “It froze the field”: many connectionists, looking back, blamed the book for steering funding and students away from neural networks for over a decade. Its reputation outran its content, as people who never read it took it to prove that neural networks were a dead end.
- “It only said what was true”: Minsky and Papert, in the 1988 expanded edition, argued that their theorems were correct and that the field had stalled for lack of new ideas, not because of their book.
- “Other forces mattered more”: historians also point to weak hardware, scarce data, Rosenblatt’s death and the overall funding squeeze. Some research — in Japan, Finland and elsewhere — never stopped.
The hard problems underneath
The deepest reasons for the winter were not reports or books but difficulties that the golden-years programs had sidestepped. Tap each to see it.
The first two have names worth remembering: combinatorial explosion and the frame problem. Both were problems of scale and openness, and both explain why microworld demos had misled everyone.
The Lighthill Report and DARPA’s cuts
In 1972 the UK Science Research Council asked Lighthill, an outsider to AI, to review the field. His report, Artificial Intelligence: A General Survey, published in 1973, sorted the work into three categories: A, advanced automation (useful engineering); C, computer-based studies of the brain and mind (respectable science); and B, the “bridge” between them, above all general-purpose robots — which he judged a failure. He singled out combinatorial explosion as the reason methods that worked on toy problems would not scale.
British support for AI research contracted to a few centres, notably Edinburgh, where Michie’s robotics work was badly hit.
In the United States the pressure came through the military budget. The Mansfield Amendment of 1969 required Defense Department research to have a clear link to military needs, which made ARPA’s open-ended “fund the people” style harder to justify. Then came a specific disappointment. ARPA’s five-year Speech Understanding Research programme (1971–1976) asked several teams to build systems that understood continuous speech from a 1,000-word vocabulary. Carnegie Mellon’s Harpy met the stated target, but only for constrained sentences in a narrow task, and ARPA — now DARPA — was unimpressed with how far that was from practical use. It did not renew the programme, and through the mid-1970s it shifted money towards projects with clear, near-term military applications.
Was the first AI winter deserved?
Choose one to hear its case.
The case
Researchers had promised translation, general problem solvers and thinking robots within years and delivered demos in microworlds. Funders are right to stop paying for promises that keep slipping. Lighthill correctly identified combinatorial explosion as a real barrier, and the cuts pushed the field towards narrower, more honest goals — which produced the useful expert systems of the 1980s.
The weak point
It treats a decade of fundamental research as if it were a failed product. Basic science routinely takes longer than expected, and some of what was cut — neural networks especially — turned out to be the right idea, merely early.
The case
The judges were outsiders applying engineering deadlines to open scientific questions. Hard problems like the frame problem had only just been named; that is progress, not failure. Abandoning neural networks and speech recognition delayed ideas that later transformed the world, and Lighthill’s three-way categorisation was itself crude.
The weak point
The field invited the backlash with its own forecasts. Had leaders been candid about how hard the problems were, the funding might have been smaller but steadier.
What survived the cold
Consensus “Winter” describes funding and public mood, not a total halt. A great deal of lasting work happened in the 1970s:
- Logic programming: Alain Colmerauer and Philippe Roussel created the Prolog language in Marseille in 1972, building on Robert Kowalski’s work in Edinburgh. It later became the flagship of Japan’s Fifth Generation project.
- Knowledge representation: Minsky’s “frames” (1974) and Roger Schank’s “scripts” offered ways to package everyday knowledge, such as what normally happens in a restaurant.
- Expert knowledge: Stanford’s DENDRAL continued, and MYCIN, which diagnosed blood infections, was built in the mid-1970s. They showed that narrow, knowledge-rich programs could be genuinely useful.
- Neural networks, quietly: Paul Werbos described a way to train multilayer networks in his 1974 Harvard thesis; Kunihiko Fukushima in Japan built the Cognitron (1975); Teuvo Kohonen in Finland and Shun-ichi Amari in Japan kept developing learning theory.
Why can a single-layer perceptron not learn XOR?
XOR is not linearly separable. A hidden layer fixes this — but training hidden layers needed backpropagation.
What did the Lighthill Report identify as a central reason AI methods would not scale?
Lighthill argued that the exponential growth of possibilities would swamp methods that worked on toy problems.
The frame problem is about…
McCarthy and Hayes (1969) asked how a logical system can capture the countless things an action does not change.
// Pause and reflect
- Lighthill was an outsider. Is an outsider better or worse placed than insiders to judge whether a research field is delivering?
- Neural networks were “early, not wrong”. How could a funder tell the difference between an idea that is early and one that is simply mistaken?
- Do you see any parallels between the forecasts that led to the 1970s winter and claims made about AI today?
Your notes are saved privately in this browser.
What happened next
The thaw came from the knowledge-rich, narrow programs that had survived. By the early 1980s companies were buying “expert systems” by the hundred and governments were launching national AI programmes — until a second, deeper chill. Continue with Expert Systems and the Second Winter.
Key terms on this page
Sources & further reading
- ALPAC, Language and Machines: Computers in Translation and Linguistics (National Academy of Sciences, 1966).
- M. Minsky & S. Papert, Perceptrons: An Introduction to Computational Geometry (MIT Press, 1969; expanded edition 1988).
- J. McCarthy & P. J. Hayes, “Some philosophical problems from the standpoint of artificial intelligence”, Machine Intelligence 4 (1969).
- J. Lighthill, “Artificial Intelligence: A General Survey”, in Artificial Intelligence: a paper symposium (Science Research Council, 1973).
- D. H. Klatt, “Review of the ARPA Speech Understanding Project”, Journal of the Acoustical Society of America (1977).
- Daniel Crevier, AI: The Tumultuous History of the Search for Artificial Intelligence (Basic Books, 1993).
- Mikel Olazaran, “A sociological study of the official history of the perceptrons controversy”, Social Studies of Science (1996).
- Lighthill report — Wikipedia overview with links to the original text and the 1973 debate.