By the late 1980s, artificial intelligence had already made a promise it could not keep.
The promise had been made in offices rather than laboratories. Systems built on rules and expert knowledge were being installed inside companies, tasked with making decisions that had once required human judgement. They were described as expert systems, and for a time they appeared to work. Investment followed. Expectations rose with it.
Then, in 1987, the ground shifted.
The systems did not fail all at once. They became expensive to maintain. Each new situation required new rules, written and verified by specialists. As the number of rules grew, so did the cost. What had been presented as automation began to resemble a different form of labour.
They also proved inflexible. An expert system could perform well within the boundaries it had been given, but outside those boundaries it faltered. It could not adapt. It could not learn from new conditions. When faced with problems that did not match its structure, it produced results that were either incorrect or unusable. In practice, this meant that systems designed for one task could not easily be extended to another.
At the same time, larger ambitions were beginning to show strain. Japan’s Fifth Generation Computer Project, launched in 1982 with the aim of developing advanced artificial intelligence systems, had drawn international attention and significant investment. By the end of the decade, it had not produced the breakthroughs that had been expected. The gap between ambition and outcome became difficult to ignore.
Economic conditions contributed to the change in mood. The stock market decline in 1987 led companies to reassess their spending. Projects that did not produce immediate and reliable returns were reduced or abandoned. Artificial intelligence, which had been promoted as a transformative technology, came to be seen as uncertain.
Alternative approaches also gained ground. Conventional software systems, based on databases and statistical methods, proved to be more predictable and less costly. They did not attempt to replicate expertise. They organised and retrieved information. For many organisations, that was sufficient.
The result was a withdrawal. Funding declined. Companies closed or moved away from artificial intelligence. Research programmes were reduced. The term itself became less common in public discussion. For a period that extended into the 1990s, the field lost much of its visibility.
This period came to be described as the second AI winter. It was not defined by a single event, but by a sustained loss of confidence. The earlier phase of optimism had been replaced by a more cautious view of what such systems could achieve.
The effect on research was uneven. Some areas contracted sharply. Others continued, though with less attention. Work on machine learning methods, including statistical models and probabilistic approaches, began to develop during this period. Neural networks, which had faced criticism earlier, were revisited, particularly following work on backpropagation in the mid 1980s. These efforts did not immediately restore the field’s reputation, but they altered its direction.
The conditions that had limited earlier systems began to change. Computing power increased. Data became more available, particularly with the expansion of the internet. Methods that relied on learning from data rather than encoding rules became more practical. By the early 2000s, these developments began to produce results that were more consistent and easier to extend.
The recovery was gradual. It did not begin with a single system or announcement. It emerged through a series of applications in search, recommendation and pattern recognition. Companies such as Google, Amazon and Microsoft incorporated these methods into services that reached large numbers of users. The systems did not present themselves as artificial intelligence. They operated within the structure of the service.
By the 2010s, advances in deep learning produced further changes, particularly in areas such as image recognition and language processing. These developments were widely described as a new phase for artificial intelligence, though they relied in part on work that had continued during the period of reduced attention.
The second AI winter is often described as a failure. It is more accurately understood as a correction. It marked the point at which expectations were adjusted to match what systems could achieve. It also shifted the focus of research away from rule based representations towards methods that could adapt to data.
The earlier systems had attempted to capture expertise directly. The systems that followed sought to infer patterns from observation. The difference between these approaches continues to shape how artificial intelligence is developed.
The period did not end with a clear boundary. Confidence returned as results improved, and as applications proved reliable. The language surrounding the field changed with it. What had once been avoided began to be used again, though with greater caution.
The collapse of the late 1980s did not prevent further progress. It altered the terms on which that progress would be judged.