drayz.ai

AI Guide · Updated 14 November 2026

History

AI winters and the rise and fall of expert systems, 1974-1993

By drayz.ai · Updated · 9 min read

Illustration of 1980s computers in a snowy landscape with a neural network sapling growing, symbolising the AI winters

Between 1974 and the early 1990s, artificial intelligence went through two "winters" of cuts and scepticism, separated by a boom in "expert systems" that briefly made AI a billion-dollar industry. This chapter explains what went wrong, what went right, and the lessons that still apply to today's AI boom.

What is an AI winter?

An AI winter is a period when funding, investment and public interest in AI fall sharply after a wave of hype fails to deliver. The term was coined at a 1984 meeting of the American Association for Artificial Intelligence, where researchers Roger Schank and Marvin Minsky warned that enthusiasm for expert systems had run ahead of reality and that a collapse would follow. They were right.

Historians usually identify two main winters:

  • The first AI winter: roughly 1974 to 1980, triggered by the Lighthill Report in the UK and funding cuts in the US.
  • The second AI winter: roughly 1987 to 1993, triggered by the collapse of the specialised AI hardware market and disappointment with expert systems.

The first AI winter, 1974-1980

The golden age of the 1960s had promised machines that would soon match human intelligence. By the early 1970s it was clear that those promises would not be kept any time soon.

In Britain, the Lighthill Report of 1973 concluded that AI research had failed to achieve its grand aims and that its methods would not scale up to real-world problems. The government cut funding for AI at most universities. In the United States, the Defense Advanced Research Projects Agency (DARPA), which had funded AI generously, began demanding research with clear military applications. A five-year DARPA programme on speech understanding ended in 1976 without producing practical systems, and DARPA cut back sharply. The neural network approach had already lost favour after Minsky and Papert's 1969 book Perceptrons.

Research did not stop. Important work continued on:

  • Knowledge representation: how to store facts and relationships so a computer can reason with them. Marvin Minsky proposed "frames" (1974), structured descriptions of stereotyped situations such as a restaurant visit; Roger Schank and Robert Abelson proposed "scripts".
  • Logic programming: Prolog, created in 1972 by Alain Colmerauer and Philippe Roussel in Marseille with theoretical work by Robert Kowalski in Edinburgh, let programmers state facts and rules and have the computer work out answers. Prolog became especially popular in Europe and Japan.
  • Expert systems in medicine and science, building on DENDRAL.

MYCIN and the knowledge revolution

The most influential expert system of the 1970s was MYCIN, developed at Stanford by Edward Shortliffe and colleagues from 1972. MYCIN advised doctors on bacterial infections and which antibiotics to prescribe. It used around 600 rules of the form "IF the infection is in the blood AND the organism is rod-shaped AND ... THEN there is suggestive evidence that the organism is ...", and it handled uncertainty with "certainty factors".

In a 1979 evaluation, a panel of experts rated MYCIN's recommendations as acceptable about as often as, or more often than, those of Stanford infectious disease specialists. Yet MYCIN was never used in routine clinical practice. The reasons are instructive and still relevant to medical AI today:

  • doctors had to type in information on slow terminals, which took too long;
  • questions of legal responsibility were unresolved;
  • clinicians were reluctant to trust a system whose reasoning they could not easily check;
  • integrating it into hospital workflows was hard.

MYCIN's lasting contribution was the insight that knowledge is power: intelligent behaviour in a narrow domain depends less on clever general reasoning methods than on capturing large amounts of specialised expert knowledge. Edward Feigenbaum called the people who extracted this knowledge from experts and turned it into rules knowledge engineers.

The expert systems boom, 1980-1987

XCON: AI makes money

In 1980 Carnegie Mellon University, led by John McDermott, delivered R1, later called XCON, to the Digital Equipment Corporation (DEC). XCON configured orders for DEC's VAX computer systems, working out which cables, cabinets and components each customer needed. It eventually had thousands of rules and was reported to save DEC tens of millions of dollars a year by reducing errors and speeding up orders.

XCON proved AI could deliver business value. Corporations rushed to build their own expert systems for credit approval, insurance underwriting, equipment fault diagnosis, oil exploration, tax advice and much more. By the mid-1980s, according to commonly cited estimates, around two-thirds of Fortune 500 companies were using or developing expert systems, and AI had become a business worth well over a billion dollars a year, including software, hardware and consulting.

Lisp machines

AI programs were usually written in LISP, which ran slowly on ordinary computers of the time. Companies such as Symbolics, Lisp Machines Inc. and Texas Instruments built specialised Lisp machines: powerful workstations designed to run LISP fast. Symbolics' machines were advanced for their day, with windowed graphical interfaces and networking, and Symbolics registered the first ever .com domain name, symbolics.com, in March 1985.

Japan's Fifth Generation project

In 1982 Japan's Ministry of International Trade and Industry launched the Fifth Generation Computer Systems project, a ten-year, government-led effort to build massively parallel computers running logic programming, which would be able to converse, translate, interpret pictures and reason like people. Japan was then an economic powerhouse that had overtaken the West in cars and consumer electronics, and the project alarmed governments in the US and Europe.

Britain's Alvey Programme

The UK responded with the Alvey Programme (1983-1987), named after John Alvey of British Telecom, who chaired the committee that proposed it. Alvey was a £350 million collaboration between government, industry and universities in advanced information technology, including "intelligent knowledge-based systems" (the British term for expert systems), very-large-scale integration chip design, software engineering and human-computer interfaces. It partly reversed the cuts that followed the Lighthill Report.

In the US, DARPA's Strategic Computing Initiative (1983) poured around a billion dollars into AI and advanced computing over the following decade, with goals such as autonomous land vehicles and a "pilot's associate". The European Union launched its ESPRIT programme.

The return of neural networks

While symbolic expert systems dominated the headlines, a smaller group of researchers revived neural networks.

  • In 1982 the physicist John Hopfield showed how a type of neural network (now called a Hopfield network) could store and retrieve memories, bringing new mathematical credibility to the field.
  • In 1985 Geoffrey Hinton and Terrence Sejnowski introduced the Boltzmann machine.
  • In 1986 David Rumelhart, Geoffrey Hinton and Ronald Williams published a landmark paper in Nature showing that backpropagation could train multi-layer networks to learn useful internal representations. Backpropagation had been described earlier, including by Paul Werbos in 1974, but the 1986 paper and the two-volume book Parallel Distributed Processing made it famous. Multi-layer networks could learn functions such as XOR that had defeated the single-layer perceptron.
  • In 1989 Yann LeCun and colleagues at Bell Labs used backpropagation to train a convolutional neural network to read handwritten postcode digits. By the late 1990s, LeCun's systems were reading a significant share of the cheques written in the United States.

Hopfield and Hinton would share the 2024 Nobel Prize in Physics for foundational work that enabled machine learning with artificial neural networks. Hinton was born in Wimbledon and studied at Cambridge and Edinburgh; much of his career was spent in Canada.

At the time, though, computers were too slow and datasets too small for neural networks to beat other methods on most tasks. Their moment would come twenty-five years later.

Robots and the "bottom-up" challenge

In robotics, Rodney Brooks at MIT challenged the whole symbolic approach. In papers such as "Intelligence Without Representation" (1991) and "Elephants Don't Play Chess" (1990), he argued that intelligence emerges from interaction with the real world, not from abstract reasoning about internal models. His behaviour-based robots, such as the six-legged Genghis (1989), used layers of simple reactive behaviours and could walk over rough ground without any central plan, much like Grey Walter's tortoises decades earlier. Brooks later co-founded iRobot, whose Roomba vacuum cleaner (2002) became the best-selling robot in history.

The second AI winter, 1987-1993

The expert systems boom ended abruptly.

The Lisp machine collapse

In 1987 the market for specialised Lisp machines collapsed. Desktop workstations from Sun Microsystems and others, and soon ordinary PCs from IBM and Apple, had become powerful enough to run AI software at a fraction of the price. An entire industry worth around half a billion dollars was wiped out within a year or two. Symbolics went into decline and eventually bankruptcy.

Expert systems hit their limits

Businesses also discovered the weaknesses of expert systems:

  • The knowledge acquisition bottleneck. Extracting rules from experts was slow and expensive. Experts often could not explain how they made decisions.
  • Brittleness. Expert systems knew nothing outside their narrow area and could fail badly on unusual cases. They had no common sense.
  • Maintenance. As rules grew into the thousands, they interacted in unpredictable ways. Updating them was costly. XCON needed a team of people to keep it current.
  • No learning. Expert systems could not improve from experience; every change had to be programmed by hand.

Grand projects disappoint

Japan's Fifth Generation project ended in 1992 having produced impressive parallel hardware and research, but none of the conversational, reasoning machines it had promised. DARPA's Strategic Computing Initiative cut AI funding sharply in the late 1980s. The UK's Alvey Programme was not renewed in its original form.

Funding dried up, AI start-ups folded and "artificial intelligence" became a label to avoid on grant applications and business plans. Researchers rebranded their work as "informatics", "knowledge-based systems", "machine learning", "computational intelligence" or "pattern recognition".

What survived the winter

Although AI's reputation suffered, much of its technology quietly lived on and spread:

  • Expert systems became business rules engines, still used today in banking, insurance, logistics and manufacturing.
  • Search and planning algorithms went into logistics, scheduling and computer games.
  • Logic programming and knowledge representation fed into databases, the semantic web and knowledge graphs, such as the ones behind Google's search results.
  • Speech recognition and handwriting recognition improved steadily using statistical methods.
  • AI planning software helped the US military schedule logistics during the 1991 Gulf War; DARPA reported that one such tool, DART, paid back its investment many times over.

The AI researcher Rodney Brooks later observed that there was a myth that AI had failed, when in fact it was all around us every second of the day; it simply stopped being called AI once it worked. This is sometimes called the AI effect: as soon as a technique becomes routine, people stop calling it artificial intelligence.

Lessons for today

The AI winters offer lessons that are directly relevant to the generative AI boom of the 2020s:

  1. Hype is dangerous. Over-promising leads to disappointment, and disappointment leads to cuts. Today's claims about imminent human-level AI may or may not prove right, but the history suggests caution.
  2. Demos are not products. SHRDLU and MYCIN were brilliant demonstrations that never became everyday tools. Integrating AI into real workflows is often the hardest part.
  3. Hardware economics matter. The Lisp machine industry died when cheaper general-purpose computers caught up. Today's AI economics depend on the price of chips, data centres and electricity.
  4. Narrow systems break. Systems that work brilliantly in a narrow, well-defined area can fail badly outside it.
  5. Maintenance costs are real. AI systems need updating, monitoring and oversight long after they are first deployed.
  6. Trust and accountability decide adoption. MYCIN failed in practice partly because of unanswered questions about responsibility and trust, the same questions facing medical and legal AI today.
  7. Good ideas can wait for their time. Neural networks were dismissed in 1969, revived in 1986 and only triumphed after 2012, when data and computing power finally caught up.

Key takeaways

  • AI winters are periods of reduced funding and interest after hype outruns results; the main ones were about 1974-1980 and 1987-1993.
  • Expert systems such as MYCIN and XCON captured human expertise as rules and briefly made AI big business in the 1980s.
  • Japan's Fifth Generation project prompted Britain's £350 million Alvey Programme.
  • Backpropagation (1986) and convolutional networks (1989) revived neural networks, though their breakthrough came much later.
  • The Lisp machine collapse and the brittleness of expert systems caused the second AI winter.
  • Much AI technology survived under other names, illustrating the "AI effect".

Previous chapter: Dartmouth and the golden age of AI · Next chapter: The machine learning era, 1993-2011