1.3 — Expert systems & MYCIN 📜

At first, researchers thought they could recreate intelligence by writing strict rules: “If this, then that.” This is called symbolic AI (or symbolic programming). The idea: lock a human expert’s knowledge into a long list of IF … THEN … rules.

🟢 In plain words — an expert system is a giant list of “if… then…” statements hand-written by humans. No learning: it blindly applies what it was told.

The iconic example: MYCIN. Developed at Stanford University in the 1970s by Edward Shortliffe, MYCIN diagnosed bacterial blood infections and recommended treatment by asking the physician a series of questions. It relied on roughly 500 to 600 rules and could even explain its reasoning when asked.

Caution

⚠️ The detail everyone misses. People often say MYCIN “didn’t work.” That’s false. In evaluations, it performed as well as the specialists in infectious diseases. And yet it was never used on real patients. Why? It couldn’t be integrated into real hospital work, plus issues of legal responsibility (“who is liable if the machine is wrong?”) and ethics. The real lesson: even an AI that works well can fail when faced with the complexity of the real world. Remember it — that’s exactly the challenge of Level 4 (going to production).

The general limitation of expert systems: the world is too complex to fit into fixed rules. As soon as a situation wasn’t anticipated in the “manual,” the system was lost. This is called the “brittleness” of these systems.

🧩 Concrete example — imagine an expert system that sorts fruit with the rule “if round and orange → orange.” Show it a tangerine (round and orange): it labels it “orange.” Show it a blood orange (reddish): it breaks, because no rule anticipated that color. That’s brittleness.

❓ Check your understanding — why did MYCIN, despite performing well, never treat real patients?

See the answer

Not for a performance reason: it diagnosed as well as the specialists. It tripped over the real world — impossible to integrate into hospital work, plus questions of legal liability (“who is at fault if the machine is wrong?”) and ethics.

✅ In short

• An expert system applies IF…THEN rules written by humans (symbolic AI).
• MYCIN (Stanford, 1970s) diagnosed infections as well as the experts.
• Its limit — and that of all symbolic AI — is brittleness when facing the unexpected.