Expert System
Encoding a specialist's knowledge as rules — AI's first commercial success, and its collapse taught the field something it's currently relearning.
When not to use it
- When the expertise is tacit. Which is most expertise. Interviewing can't extract what the expert can't access.
- In an open domain. Rules cover what you wrote; reality doesn't stop, and these fail catastrophically rather than gracefully.
- At scale. Rules interact. Knowledge bases become unmaintainable at the size where they become useful.
- When you have data. Learning from examples sidesteps the whole bottleneck. That's why this generation works.
Reach for something else instead
- Learning from data — the answer to Polanyi's paradox. Learn the tacit thing from artefacts.
- LLM + tools — the modern shape, with the modern version of the same trap.
- Decision trees learned from data — interpretable and not hand-authored.
- Business rules engines — where rules are genuinely the spec (tax, compliance), this still works fine.
Sources & further reading
- Buchanan & Shortliffe (1984), Rule-Based Expert Systems: The MYCIN Experiments — the system that beat the faculty and never shipped.
- Feigenbaum (1977), The Art of Artificial Intelligence — the knowledge-is-power thesis, from the field's founder.
- Polanyi (1966), The Tacit Dimension — "we know more than we can tell." The reason it was never going to work.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Concluding the lesson was "symbolic AI failed." It was "hand-authored knowledge doesn't scale," and the syntax is irrelevant.
- Missing that MYCIN worked. It outperformed the faculty and died on liability and workflow.
- Building 3,000-line system prompts and not noticing you've built a knowledge base with all the same properties.
- Forgetting they had real explainability. The rules were the reasoning. We traded that for capability.
At a glance
FieldFoundations
The 1980sAI's first real industry, billions spent
MYCINoutperformed Stanford's infectious disease faculty, never deployed
What killed itthe knowledge acquisition bottleneck, and Polanyi's paradox underneath it
What it had that we don'treal explanation; the rules were the reasoning
The rhymea 3,000-line system prompt is a knowledge base
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Expert system vs. LLM — one asked experts to dictate what they know and hit Polanyi's paradox; the other learned the tacit part from what experts produced. That's the whole difference.