AI's history has been marked by cycles of hype and disappointment. When promises went unfulfilled, funding collapsed, a period researchers call an AI winter. Between winters, new approaches brought fresh excitement and money.
Machine translation stalls
One early disappointment concerned translation. During the Cold War, the US government funded research on translating Russian into English by computer. A famous 1954 demonstration by Georgetown University and IBM translated about sixty carefully chosen sentences, and researchers predicted that translation would be solved within a few years. It was not. Words have many meanings, grammar is complex, and translation requires understanding context. In 1966, a US government committee issued the ALPAC report, concluding that machine translation was slower, less accurate and more expensive than human translation, and funding was sharply cut.
The first AI winter
In Britain, the government asked the mathematician Sir James Lighthill to evaluate AI research. His 1973 report concluded that AI had failed to achieve its grand objectives, and that its methods worked only on toy problems and broke down when faced with the combinatorial explosion: the number of possibilities a program must consider grows so fast in real problems that no computer can search them all. The British government cut AI funding at most universities. In the United States, the defense research agency DARPA also cut back on open-ended AI research in the mid-1970s, demanding more practical results. The period from roughly 1974 to 1980 is often called the first AI winter.
Expert systems
AI revived in the late 1970s and 1980s with a new idea: instead of trying to build general intelligence, capture the specialized knowledge of human experts in a narrow field. These programs, called expert systems, consisted of a knowledge base of if-then rules, gathered by interviewing experts, and an inference engine that applied the rules to specific cases.
At Stanford, DENDRAL, begun in 1965, helped chemists identify molecules from mass spectrometry data. MYCIN, developed in the 1970s, diagnosed bacterial blood infections and recommended antibiotics. In a 1979 evaluation, its recommendations were judged as good as or better than those of infectious disease experts. Yet MYCIN was never used in hospitals, because of concerns about legal responsibility, the difficulty of integrating it into medical practice, and the slowness of entering data on computers of the time.
The boom
Expert systems became commercially successful. In 1980, the Digital Equipment Corporation began using a system called XCON (originally R1) to configure orders for its computers, checking that each customer received the right combination of components. The company reported savings of tens of millions of dollars a year. Corporations across the world built expert systems, and a new industry grew up, including companies making specialized Lisp machines designed to run AI software.
In 1982, Japan's Ministry of International Trade and Industry launched the Fifth Generation Computer Systems project, a ten-year national effort to build computers based on logic programming that could reason and converse. Alarmed, the United States and Britain launched their own large programs. AI seemed poised to transform industry.
The second winter
By the late 1980s, the boom collapsed. Expert systems proved brittle: they worked within their narrow domain but failed badly when faced with unexpected cases, and they could not learn. Keeping their thousands of rules up to date was expensive and difficult, a problem called the knowledge acquisition bottleneck. In 1987, the market for Lisp machines collapsed as cheaper desktop computers from companies such as Apple and IBM became powerful enough to run the same software. Japan's Fifth Generation project ended in 1992 without reaching its main goals. Funding fell again, and many AI companies failed. The period from about 1987 to the early 1990s is called the second AI winter. Many researchers avoided the term “artificial intelligence,” describing their work instead as informatics, knowledge systems or machine learning.
Common sense
One ambitious project tried to tackle the common sense problem directly. In 1984, the researcher Douglas Lenat began Cyc, an effort to hand-write millions of facts and rules about everyday life, such as that water makes things wet and people cannot be in two places at once. The project continued for decades and produced useful tools, but it never achieved general common sense reasoning, illustrating how vast and subtle everyday knowledge is.
Worked example: anatomy of a hype cycle
Trace the pattern that repeated in AI's history, using expert systems as the case.
- Breakthrough. A real advance, such as MYCIN or XCON, shows impressive results in a narrow task.
- Inflated expectations. Investors, governments and journalists assume the approach will soon handle much broader tasks.
- Overinvestment. Money floods in; many companies form; grand national projects launch.
- Disappointment. Limits appear: brittleness, cost, difficulty scaling.
- Winter. Funding collapses, and even genuinely useful work is tarred with failure.
- Quiet survival. Useful parts continue under other names. Rule-based systems are still used in business software today.
The pattern is not unique to AI; it appears in railways, the internet and many other technologies. Recognizing it helps in judging any new technology: the right question is rarely whether it is magic or useless, but which specific tasks it can reliably do, at what cost.