Expert Systems: The AI Winter They Built

In the 1970s, AI researchers shifted away from general problem-solving toward domain-specific expertise. The insight: instead of building general intelligence, encode the knowledge of a specific expert, a cardiologist, a geologist, a tax lawyer, into a rule-based system. The result was the expert system: a program that reasoned through chains of IF-THEN rules drawn from interviews with human experts.

MYCIN, developed at Stanford between 1972 and 1980, was the most successful early example. It diagnosed bacterial blood infections and recommended antibiotic treatments, matching or exceeding the accuracy of specialists in controlled studies, around 65% accuracy versus 42–62% for human physicians. It explained its reasoning step-by-step. Clinically, it worked. And yet MYCIN was never actually used in clinical practice. The barriers were legal liability, resistance from doctors, and the difficulty of keeping its 600 rules current as medical knowledge evolved.

Through the late 1970s and 1980s, expert systems became a commercial boom. Companies poured hundreds of millions of dollars into AI shells, software frameworks for building expert systems. Banks, oil companies, and manufacturers deployed rule-based systems for loan decisions, geological surveys, and equipment diagnosis. The commercial AI industry briefly outpaced academic research. Lisp machines, specialised hardware for running AI code, became a multi-hundred-million-dollar market.

Then, in the late 1980s, the market collapsed. The maintenance cost of rule bases was staggering, as domains changed, rules needed constant updates, and the expertise to maintain them was expensive. Lisp machine manufacturers were wiped out by cheaper general-purpose workstations. DARPA cut funding. The second AI winter arrived. Expert systems had been enormously useful in narrow domains, but they could not generalise, could not learn, and could not adapt without human intervention. Neural networks, quietly developing in universities, were waiting.