Between 1980 and 1987, artificial intelligence was not a research curiosity but a commercial juggernaut. Expert systems—software that encoded human expertise as thousands of if-then rules—attracted more than a billion dollars in venture capital and corporate investment. Companies like Teknowledge, IntelliCorp, and Carnegie Group went public. Digital Equipment Corporation deployed XCON, a system with 2,500 rules for configuring VAX computers, and claimed it saved thirty million dollars annually. The promise was seductive: capture the knowledge of your best engineer, cardiologist, or geologist in software, then replicate that expertise infinitely.
The architecture was elegant in its simplicity. A knowledge engineer would interview domain experts, distill their heuristics into formal rules, and load them into an inference engine. MYCIN, developed at Stanford in the mid-1970s, diagnosed bacterial infections with about 450 rules and reportedly matched the accuracy of human specialists. DENDRAL identified molecular structures from mass spectrometry data. The systems worked, within their narrow domains, and the AI community believed it had found the path to machine intelligence.
The brittleness problem
Expert systems failed not because the logic was wrong but because it could not scale. Every new scenario required new rules, hand-coded by knowledge engineers who became bottlenecks. XCON eventually grew to more than 10,000 rules and required a full-time team just to maintain it as hardware configurations changed. The systems were brittle: ask a question slightly outside their training and they collapsed. They had no common sense, no ability to learn from examples, and no graceful degradation. Worse, extracting knowledge from human experts proved far harder than anticipated—experts often could not articulate the intuitions that guided their decisions.
By 1987 the market had evaporated. Hardware companies that sold specialized LISP machines went bankrupt. The AI winter set in, funding dried up, and the term "artificial intelligence" became toxic in boardrooms for more than a decade. The collapse was so complete that an entire generation of computer scientists learned to avoid the phrase.
What neural networks learned
The failure of expert systems created the intellectual space for connectionism to return. Researchers like Geoffrey Hinton and Yann LeCun, working on neural networks in relative obscurity during the 1980s, offered a fundamentally different approach: instead of encoding knowledge as rules, learn patterns from data. The contrast was stark. Expert systems required exhaustive manual labor and produced fragile, narrow tools. Neural networks required large datasets and computing power but could generalize, handle ambiguity, and improve with scale.
Every major AI breakthrough since—ImageNet, AlphaGo, GPT, diffusion models—rests on the lesson that intelligence emerges from learning, not from rule-writing. The transformer architecture that powers modern language models has no hand-coded heuristics about grammar or meaning; it discovers structure by predicting the next token across trillions of words. The irony is that expert systems were not wrong about their goal—capturing and scaling expertise—but about their method. Modern AI achieves what MYCIN and XCON promised, but through statistical learning rather than symbolic logic.
Our take
The expert systems era is remembered as a cautionary tale, but it deserves more credit. It proved that businesses would pay for AI if it solved real problems, established the economic model for enterprise software, and trained a generation of engineers who later built the internet economy. Its failure was necessary: AI had to learn that human knowledge is not a list of rules but a web of learned associations, and that the only way to capture it is to let machines learn the way humans do—through experience, not instruction. The winter that followed cleared out the hype and left space for the patient, empirical work that eventually succeeded.
