What actually caused the AI winters

Marvin Minsky and Frank Rosenblatt went to the same high school, a year apart. One became the leading advocate of symbolic AI, the other invented the perceptron and became the leading advocate of learning from data.

In 1969 Minsky and Seymour Papert published a book that is widely credited with ending neural network research for a decade. Rosenblatt died in a boating accident in 1971. When the expanded edition appeared in 1987, correcting several errors in the original, it carried a handwritten dedication to him.

That is the human shape of the story. The intellectual shape is different from the one the field repeats, and it matters because the version people tell has a moral attached: a famous researcher was wrong, and the field lost a decade to his mistake.

The theorem in that book was correct. The conjecture attached to it was wrong. The field's reception conflated them, and the conflation did the damage rather than the error. That pattern, where a narrow true result is received as a broad false one, recurs through both winters and is the actual lesson.

Myth one: Perceptrons was wrong

The book proved something specific and true. A single-layer perceptron cannot learn a function that is not linearly separable, and XOR is the standard example. No arrangement of weights on one layer separates the XOR cases, and the proof is elementary and permanent.

That is a real limitation of a real architecture, and stating it was correct scholarship rather than an attack.

What was wrong was a conjecture the authors added: that comparable limits would apply to multilayer networks. They noted such networks could represent XOR and observed that no training algorithm for them was known. Then they speculated the extension would fail.

It did not. Multilayer networks learn XOR, and Grossberg published networks modelling it within three years. Backpropagation supplied the missing training algorithm and the objection evaporated.

So the record is: a true theorem about one-layer networks, plus a false guess about many-layer networks, published together. The field took away "neural networks are limited" without the qualifier that made the claim true, which is a failure of reception rather than of authorship, which is the same shape as the vacuous-bound confusion. The 1987 edition corrected errors, and by then the damage was cultural rather than technical.

Myth two: the Lighthill Report caused the first winter

The 1973 report is routinely described as the trigger. It was highly critical, it concluded AI had failed to achieve its stated objectives, and it was followed by the UK effectively ending university AI funding, with researchers leaving the field or the country and British AI not recovering its position for decades.

All of that is accurate and it makes the report look like a cause. The dates say otherwise.

The ALPAC report cut machine-translation funding in 1966, three years before Perceptrons and seven before Lighthill. Confidence was already eroding, DARPA was already shifting from open-ended research toward directed work with visible applications, and the money was already moving.

Lighthill formalised a retreat that was underway. Reports of that kind do not usually initiate a change in funding mood; they are commissioned because the mood has already changed and someone wants it documented.

What Lighthill actually argued, which was right

The report is remembered as a hatchet job, and its central technical claim was correct and remains correct.

Lighthill argued that AI methods faced combinatorial explosion: the number of possibilities to examine grows exponentially with problem size, so techniques demonstrated on small examples become computationally impossible at realistic scale. Search over a toy world is tractable; the same search over a real one is not.

That was true in 1973. It is still true, and it is why every serious system since has been an attempt to avoid exhaustive search rather than to speed it up. Heuristics, learned value functions, pruning, statistical methods and neural approximation are all responses to the same problem Lighthill named.

The report was wrong about prospects and right about the obstacle, which is a more useful thing to remember than "it was unfair."

The overpromising came from inside the field

Convenient to forget, and load-bearing.

In 1970 Minsky said publicly that within three to eight years there would be a machine with the general intelligence of an average human being. That was not journalism or a marketing department. It came from the most prominent researcher in the field, and predictions of that kind shaped both funding decisions and public expectation.

When the expectation was not met, the correction was not proportionate to the actual scientific progress, which had been real. It was proportionate to the gap between promise and delivery, and the promise had been set by researchers themselves.

Winters are not caused by systems failing. They are caused by systems failing to match claims, and the claims were mostly made by people who should have known the uncertainty.

The second winter was a hardware market collapse

Underappreciated, and it changes what the episode teaches.

Expert systems revived the field in the early 1980s. They worked, delivered commercial value in narrow domains like medical diagnosis and financial analysis, and by 1985 corporations were spending over a billion dollars annually on AI, most of it internal. An industry formed around them: software firms and, critically, hardware firms selling specialised LISP machines built to run the systems.

Then in 1987 that market collapsed.

What failed was the specialised hardware, not the software. General-purpose workstations from Sun and others became fast enough and cheap enough to run the same systems, so the case for a dedicated machine costing many times more disappeared. The companies died. The expert systems kept running, many of them for years, on ordinary computers.

The winter that followed is remembered as evidence that expert systems did not work. They worked. A hardware business model built around them did not, and the field drew a broader conclusion than the evidence supported. Same failure mode as the perceptron reception: a narrow result received as a general one.

The dates, laid out

The narrative version compresses events that were years apart, which is how a confirmation gets remembered as a cause. Set against a timeline it reads differently.

1958 Rosenblatt demonstrates the perceptron. Coverage is extravagant. 1966 ALPAC reports on machine translation. Funding for it is cut. This is the first retreat and it precedes everything usually blamed. 1969 Perceptrons published. A true theorem about single layers, a false conjecture about multiple. 1970 Minsky predicts human-level general intelligence within three to eight years. 1971 Rosenblatt dies. 1972 Grossberg publishes networks modelling XOR, three years after the conjecture. 1973 Lighthill reports. UK university funding effectively ends. 1974 DARPA cuts deeply. The first winter is under way. 1980s Expert systems revive the field commercially. 1985 Corporate AI spending exceeds a billion dollars annually. 1986 Backpropagation is popularised, supplying the training algorithm whose absence the 1969 conjecture rested on. 1987 The specialised LISP machine market collapses. Perceptrons is reissued with corrections and a dedication to Rosenblatt.

Two things stand out once it is sequenced.

The counter-evidence to the conjecture arrived in 1972, a year before Lighthill and two before the funding collapse. It was available throughout the first winter and did not change the reception, which indicates the problem was never that the field lacked the correction.

And the first funding cut came in 1966, before the book everyone blames. The causal story runs backwards from where it is usually placed.

What actually produces a winter

Assembling the record, the pattern has three components and the middle one is doing most of the work.

Real capability, narrower than described. Perceptrons truly learned. Expert systems truly encoded expertise. Neither was fraudulent and neither was as general as its advocates suggested.

A gap between the claim and the delivery, opened by the field itself. This is the necessary ingredient. Slow progress alone does not cause a funding collapse; slow progress against a promised timeline does.

A precipitating document or event that formalises what people already suspect. ALPAC, Lighthill, the 1987 hardware collapse. These are remembered as causes and function as confirmations.

The component that is absent from the list is worth noticing: in neither winter did the underlying technology stop working. Perceptrons still learned linearly separable functions. Expert systems still ran. What changed was the willingness to fund the difference between what existed and what had been described.

Does the pattern apply now

The honest answer is that two of three components are visible and the third is not, and the third is the one that matters.

Real capability narrower than described: clearly present. Systems that perform impressively on demonstrations and unevenly in production, marketed with claims that outrun evaluation.

A gap opened by the field itself: also present, and the timeline predictions are being made by the same class of people who made them in 1970.

A precipitating event: not yet. And the structural difference is that both previous winters occurred when AI was funded almost entirely by government research budgets, which can be cut by a small number of decisions. Current funding is commercial and tied to revenue, which behaves differently: it does not vanish on a report, it erodes when customers stop renewing.

That may make a sharp winter less likely and a long disappointment more likely, which is a different failure mode and possibly a worse one, because a winter ends.

What is unresolved

Whether the winters were net harmful. Both redirected funding toward narrower, more accountable work, and some of what emerged, including backpropagation in the mid-1980s, came from researchers working through the cold period. Whether the field would have arrived faster without the interruptions is unanswerable.

How much the Perceptrons reception actually delayed connectionism. The standard claim is a lost decade. Work continued throughout, the missing piece was a training algorithm rather than permission, and whether that algorithm would have arrived sooner under different social conditions is not knowable.

Whether commercial funding is more or less stable than government funding. It is more diversified and more directly tied to whether the thing works, and it is also faster to move. Nobody has a good model of how an AI retreat would propagate through a commercially funded field, because it has not happened.

And whether the term itself is useful. Recent work has questioned whether the standard explanations of overpromising, hardware limits and brittleness capture what happened, or whether something more structural about paradigm fragility is involved. Treating the winters as a single repeatable phenomenon may be a convenience rather than a finding.

The counter-argument

The myth-correction can be overdone. Minsky and Papert were not neutral parties writing a technical note. They were the leading advocates of a competing research programme, publishing a critical book about a rival approach, and the conjecture about multilayer networks was not an incidental error but the part that made the book consequential. Emphasising that the theorem was correct risks understating that.

The precipitating documents may matter more than this account allows. Institutional decisions need justification, and a report from an eminent scientist provides one. It is possible that without Lighthill, funding would have declined more gradually and the field would have kept more people.

And the comparison to now may be reassurance dressed as analysis. Every generation has believed its situation was structurally different from the previous downturn. The observation that current funding is commercial rather than governmental is true and does not establish it is more robust, since commercial funding has its own history of abrupt reversals.

The short version

Minsky and Rosenblatt attended the same high school a year apart and became the leading advocates of opposing approaches. The 1969 book credited with ending neural network research proved a true theorem, that a single-layer perceptron cannot learn a function that is not linearly separable, and attached a false conjecture, that comparable limits would apply to multilayer networks. Multilayer networks learn XOR, Grossberg demonstrated it within three years, and backpropagation supplied the missing training algorithm. The field remembered the conjecture and dropped the qualifier that made the claim true.

The Lighthill Report of 1973 is remembered as the trigger for the first winter and functioned as a confirmation. ALPAC had already cut machine-translation funding in 1966, three years before Perceptrons and seven before Lighthill, and DARPA was already shifting toward directed research. What Lighthill actually argued was that AI faced combinatorial explosion, where possibilities grow exponentially with problem size, and that claim was correct then and remains correct now: every serious system since has been an attempt to avoid exhaustive search rather than accelerate it.

The overpromising came from inside. Minsky predicted publicly in 1970 that a machine with average human general intelligence was three to eight years away. Winters are not caused by systems failing but by systems failing to match claims, and the claims were made by people positioned to know the uncertainty.

The second winter was a hardware market collapse rather than a capability failure. Expert systems worked and delivered value, corporations were spending over a billion dollars annually by 1985, and in 1987 the market for specialised LISP machines collapsed because general-purpose workstations became fast and cheap enough to run the same software. The systems kept running. The field concluded expert systems had failed.

Both episodes share one shape: a narrow true result received as a broad false one. And in neither winter did the technology stop working. What collapsed was the willingness to fund the difference between what existed and what had been described.

Common questions

What was the AI winter? A period of sharply reduced funding and interest in artificial intelligence following a cycle of high expectation and disappointment. Two are conventionally identified: roughly 1974 to 1980, after the critical reception of symbolic and perceptron-based programmes, and roughly 1987 to 1993, after the collapse of the expert-systems hardware market. In neither case did the underlying technology stop working; funding for the gap between claims and delivery is what disappeared.

What did Minsky and Papert actually prove? That a single-layer perceptron cannot learn a function that is not linearly separable, with XOR as the standard example. The proof is elementary and permanent, and it is a correct statement about that architecture. What was wrong was a separate conjecture in the same book, that comparable limitations would extend to multilayer networks. They do not, and the authors themselves noted multilayer networks could represent XOR while no training algorithm for them was then known.

Did Perceptrons kill neural network research? It contributed, and the mechanism is usually described inaccurately. The book's central theorem was true; the damage came from the field receiving "neural networks are limited" without the qualifier restricting it to a single layer. Work continued through the period, the missing piece was a training algorithm rather than permission, and backpropagation supplied it in the mid-1980s. A 1987 expanded edition corrected several errors and carried a dedication to Rosenblatt, who had died in 1971.

What was the Lighthill Report? A 1973 survey of AI research for the British Science Research Council, concluding that the field had not produced the impact promised. It was followed by the UK effectively ending university AI funding, with lasting damage to British research. It is remembered as the cause of the first winter and is better read as a confirmation, since ALPAC had already cut machine-translation funding in 1966 and DARPA was already moving toward directed research.

What is combinatorial explosion? The problem that the number of possibilities a search must examine grows exponentially with problem size, so methods that work on small examples become computationally impossible at realistic scale. It was Lighthill's central technical argument in 1973 and it was correct. Every subsequent approach, from heuristics and pruning to learned value functions and neural approximation, is an attempt to avoid exhaustive search rather than to perform it faster.

What caused the second AI winter? A hardware market collapse rather than a failure of the software. Expert systems worked and delivered commercial value in narrow domains, and by 1985 corporations were spending over a billion dollars annually on AI. In 1987 the market for specialised LISP machines collapsed, because general-purpose workstations had become fast and cheap enough to run the same systems for far less. The companies selling dedicated hardware died; the expert systems continued running on ordinary computers.

Could another AI winter happen? Two of the three historical components are present: capability narrower than described, and a gap between claims and delivery opened by the field itself. The third, a precipitating event, is not. The structural difference is that both previous winters occurred when AI was funded almost entirely by government research budgets, which can be cut by a small number of decisions, whereas current funding is commercial and tied to revenue. That may make a sharp winter less likely and a long disappointment more likely, which is a different failure mode and arguably worse, since a winter ends.

What is the actual lesson from the AI winters? That a narrow true result can be received as a broad false one, and the reception does more damage than any error. It happened to the perceptron theorem, which was correct about one layer and remembered as a verdict on neural networks generally, and to the expert-systems collapse, which killed a hardware business model and was remembered as evidence the software did not work. Both times the technology kept working and the willingness to fund the gap between claim and delivery is what collapsed.