The machine was right. The room wasn't ready to believe it.
In 1979, a panel of Stanford physicians sat down to grade a stack of anonymized treatment recommendations for dangerous blood infections. They didn't know that one of the "doctors" in the stack wasn't a doctor at all. It was a program running on a mainframe, built mostly by a graduate student. When the scores came back, the program's recommendations held up against the specialists — in some categories, it outscored them. And then almost nothing happened. No hospital adopted it. No patient was ever treated by it. It went back in a drawer.
The Grad Student Nobody Remembers
Episode 7 tells you MYCIN was built at Stanford by Edward Feigenbaum and Bruce Buchanan. That's true, but it undersells the story. The person who actually sat down and wrote the thing, rule by rule, was a young MD-turned-computer-scientist named Ted Shortliffe, working it into his PhD dissertation under Buchanan's and Feigenbaum's supervision. Shortliffe spent years doing something closer to journalism than programming — sitting across from infectious disease specialists, asking them to explain decisions they'd made so instinctively they'd never had to put words to them before. That interviewing process is the part the field would later admit was the real bottleneck, and Shortliffe was doing it years before anyone had a name for the problem.
Right, and Shelved Anyway
The 1979 evaluation, run by a team including Victor Yu, is one of the strangest results in AI history precisely because it wasn't a story about failure. MYCIN worked. So why the drawer? Partly liability — nobody could answer who gets sued if a machine's prescription goes wrong. Partly workflow — MYCIN expected a doctor to sit at a terminal answering dozens of questions, which is not how a hospital floor operates. But underneath both was something more human: physicians were being asked to defer to a machine's judgment on decisions that were, quite literally, life and death, and the profession wasn't built to metabolize that kind of trust transfer. The tacit-knowledge problem the field spent the next decade fighting had already shown up here, on the other side of the interaction — not just in what the machine couldn't learn, but in what humans wouldn't hand over.
Why This Still Matters
That tension — a system performing well and still being sidelined — is the version of the Expert Systems story that never made it into the technical retrospectives. It's the same tension showing up today in every argument about whether an AI tool that "tests well" is actually ready to be trusted with a real decision. Performance and adoption have never been the same question.
The video picks up the technical half of this story — the architecture, the bottleneck, the assumption baked into the rules that eventually broke the whole approach. Watch it and you'll understand exactly why a system that could match a specialist still couldn't survive contact with the real world.
