The Expert Systems Era

Imagine asking an experienced doctor to write down every clue behind every decision. The obvious rules come first. Then arrive the exceptions, interactions, rare cases, and judgements the doctor uses without being able to state every step.

For a while in the 1980s, that was the plan, and it was not a bad one. Sit with the experts, get the knowledge out of them and onto paper, and hand the paper to a machine. It worked. There were systems doing real medicine, narrow and genuinely useful, and every step of their reasoning could be read and argued with by a specialist. That last part is worth holding onto, because almost nothing built since can offer it.

Then comes the failure, and its shape is the interesting thing. The rulebook does not break loudly. It has no way to know that this case is one nobody anticipated, so it does what it always does: matches, concludes, and answers with total confidence. Being wrong feels exactly the same to it as being right.

You can patch that, one exception at a time, and people did, for years. The trouble is that the exceptions do not run out, and each patch can disturb something that already worked.

There is a lesson buried in the collapse that followed, and it is not that the idea was stupid. It is that a thing can work exactly as designed and still be sold as something it was never going to be.