AI Winter 1974-1993: The Promise and the Silence

Artificial intelligence lost funding and public trust for nearly two decades, from 1974 to around 1993. The cause was not broken machines, but promises that outpaced what the technology could deliver. The pattern repeated twice, and it was almost exactly the same both times.

Both began with demos that looked too impressive for their cost. In 1966, ELIZA by Joseph Weizenbaum only matched keywords, then flipped sentences into questions. Many users still believed the program had feelings. In 1970, SHRDLU by Terry Winograd could chat about blocks on a virtual table. Both were presented as major advances, but the substance was far thinner.

1966 and 1970: Demos That Looked Too Smart

ELIZA worked like a shopkeeper who memorized one message, then said it to everyone who walked by. What shocked people was not the technology, but how little it took to fool them. One small script was enough to make people feel they were talking to something that understood.

SHRDLU was more convincing. It could understand commands like "put the red block on the blue block" in a virtual block world. But its world was one table, one type of object, and one set of rules. Outside of that, nothing carried over. This was the pattern that would repeat for two decades: extraordinary ability inside a box, and zero outside it.

1973: The Report That Changed Everything

In 1973, Sir James Lighthill wrote a report for the UK Science Research Council. Its message was blunt: most AI research had not shown progress proportional to the funding it received. His recommendation was clear — cut support.

That report was not the only pressure. DARPA, the largest US funder, began demanding evidence that AI projects produced something usable by the military. Many projects failed to meet that standard. By 1974, AI funding on both sides of the Atlantic had dropped sharply.

What is interesting is that Lighthill was not wrong about the technology. What was wrong was how he measured success. He judged AI by whether it could already do difficult tasks, not by whether it could do useful ones. That difference is small on paper, but large in the lab.

1969: The Book That Got Blamed

Before Lighthill, there was "Perceptrons" by Marvin Minsky and Seymour Papert, published in 1969. The book proved mathematically that single-layer perceptrons cannot learn the XOR function. The conclusion was correct, but it was too often read as "neural networks are useless."

Modern historians debate how much the book actually caused the AI winter. What is clear is that it gave academics a reason to pull funding from neural network research. And that was enough. A book can stop a field for fifteen years with a single misunderstood theorem.

1980s: The Second Bubble

Funding rose again in the early 1980s. This time it was not neural networks, but expert systems — programs that mimic how experts use if-then rules. XCON at Digital Equipment Corporation saved millions by configuring VAX computers. MYCIN at Stanford diagnosed blood infections with manually written rules.

Japan joined with the Fifth Generation Computer Project in 1982, with a budget of hundreds of millions of dollars. The goal was to build computers that could think like humans within a decade. The target was too optimistic, and by the end of that decade the project was considered a failure.

In the market, LISP machines became the symbol of the era. LISP was a programming language built for AI, and LISP machines were specialized computers that ran it fast. Companies like Symbolics and Lisp Machine Inc. sold these machines to universities and laboratories for tens of thousands of dollars.

1987: The Market Collapses

LISP machines needed specialized hardware. When general-purpose workstations from Sun and DEC became faster and far cheaper, the reason to buy LISP machines disappeared. By 1987, that market had collapsed. Symbolics and Lisp Machine Inc. began to face financial trouble.

Expert systems also began to show their limits. Manually written if-then rules could not handle unexpected exceptions. XCON was great for standard configurations, but failed when a request was slightly different. The cost of maintaining ever-growing rules eventually exceeded the benefit.

The Pattern That Repeats

Both winters had the same pattern. First, an impressive demo made expectations explode. Second, those expectations could not be met within the promised timeframe. Third, funding was pulled, and research moved to lower-profile places.

What is interesting is that the technology did not die. Neural networks continued to be developed in small laboratories throughout the 1980s. Backpropagation, the algorithm that became the foundation of modern deep learning, was published in 1986 by Rumelhart, Hinton, and Williams — right in the middle of the winter. Nothing stopped them, because no one considered them important.

Analysis: Why This Still Matters

In 2026, the same pattern is appearing again. Large language models produce impressive demos. Public expectations explode. The question is not whether the pattern will repeat, but when.

The lesson is simple: a demo is not proof. Lasting progress needs more than one stunning demonstration. It needs consistent capability, reasonable cost, and applications that people actually use.

What is most often forgotten is that the AI winter did not kill the technology. It only moved it. Neural networks did not disappear in 1974. They moved to small laboratories, kept being developed by people who did not need public attention to work. When the time came, the technology was ready. What was not ready was public expectation, which had to be built from scratch again.

What changed between the two eras was not the technology, but how people measured progress. In the 1970s, the measure was whether a machine could do a difficult task. In the 1980s, the measure was whether a machine could replace a human. Both were too ambitious for their time, and both paid the same price.

The hardest part of this pattern is that it does not feel like a pattern while it is happening. Every era is certain that this time is different, that the technology is mature enough, that this time it will not fail. That belief is not a sign of failure. It is a prerequisite for funding to flow in. Without that belief, no one writes the check.

What Did Not Change

Three things did not change from 1974 to 2026. First, public expectations always move faster than real capability. Second, funding follows expectations, not capability. Third, abandoned technology often becomes the foundation for the next generation.

Neural networks were abandoned in 1974, then became the foundation of deep learning in 2012. Backpropagation was published in 1986, then became standard in the 2010s. This pattern will not stop. What can be done is recognize it sooner.

The question is not whether another AI winter will come. The question is whether we will recognize the pattern before it is too late. And whether we will have the discipline to act on that recognition, even when the demos are still impressive and the funding is still flowing.

The hardest part of this pattern is that it does not feel like a pattern while it is happening. Every era is certain that this time is different, that the technology is mature enough, that this time it will not fail. That belief is not a sign of failure. It is a prerequisite for funding to flow in.