Expert Systems: AI Goes to Work

By the mid-1970s, AI researchers pivoted from general intelligence to narrow expertise. The bet was that if you could interview a specialist, extract their decision rules, and encode them as if-then statements, the resulting program would match the expert on a well-defined task.

Edward Feigenbaum at Stanford led this shift. DENDRAL analyzed chemical mass spectra. MYCIN diagnosed bacterial infections and recommended antibiotics, reportedly outperforming some junior doctors in trials. XCON helped Digital Equipment Corporation configure VAX computers and saved the company tens of millions of dollars a year.

Anatomy of an expert system

• A knowledge base of rules elicited from human experts
• An inference engine that chains rules forward or backward
• A working memory of current facts about the case
• An explanation module that justifies conclusions in English

The field spawned an industry. Companies sold shells like KEE and ART, and knowledge engineer became a paid profession. Japan announced its Fifth Generation Computer Project in 1982, betting billions on parallel logic machines to dominate the coming AI era.

“In the knowledge lies the power.”

The big idea: symbolic AI genuinely delivered value for narrow, stable domains. Its collapse came not from failure but from the impossibility of scaling hand-written rules to everything humans know.