The Rise of Expert Systems
Expert systems were computer programs built in the 1970s and 1980s that encoded the knowledge of human specialists — doctors, geologists, engineers — into a set of logical rules, allowing machines to mimic expert decision-making for the first time. They matter because they represent the first serious commercial success of artificial intelligence and established the idea that machines could reason, not just calculate. The counterintuitive fact: these systems had no learning ability whatsoever — every piece of "intelligence" had to be painstakingly typed in by hand, yet they still outperformed many human experts in narrow domains.
The "aha moment" for expert systems is this: intelligence, at least in a narrow domain, can be written down as a list of rules. If a doctor diagnoses pneumonia by checking for fever, cough, and a specific X-ray pattern, you can literally type "IF fever AND cough AND abnormal X-ray THEN suspect pneumonia" — and a computer will follow that logic flawlessly, every time, without getting tired or distracted. Expert systems proved that knowledge itself was the engine of intelligence, not raw computing power or mysterious human intuition.
How It Works
At the heart of every expert system are three components working together like a well-organized office. First, there is the knowledge base — a library of facts and rules provided by human experts. Second, there is the inference engine — the reasoning machine that reads those rules and applies them to a specific problem. Third, there is the working memory — a scratchpad where the system keeps track of what it currently knows about the situation in front of it. The rules themselves follow a simple pattern called IF-THEN logic. An example from the famous medical expert system MYCIN might look like: > IF the organism is gram-negative AND the organism is a rod AND the patient is a compromised host THEN there is evidence that the organism is Pseudomonas aeruginosa. The inference engine reads the current facts in working memory, scans through all the rules, finds every rule whose IF part matches the known facts, and then "fires" that rule — adding its conclusion to working memory. This process repeats, with new conclusions becoming new facts that trigger more rules, cascading forward until the system reaches a final answer. This chain reaction of reasoning is called forward chaining. Alternatively, some systems worked backwards — starting with a hypothesis ("Could this be meningitis?") and then asking questions to gather only the facts needed to confirm or deny it. This is called backward chaining, and it was much more efficient when the goal was already known.
Edward Feigenbaum studied under Herbert Simon at Carnegie Mellon before moving to Stanford, where he led the Heuristic Programming Project and built some of the earliest and most influential expert systems. His key contribution was the insight that the bottleneck in AI was not processing power but knowledge — what he called the "knowledge is power" hypothesis — which he demonstrated through systems like DENDRAL (chemistry) and MYCIN (medicine). Surprisingly, Feigenbaum won the Turing Award in 1994 partly for work that the broader AI community had already begun to dismiss as a dead end, a sign of how hotly contested expert systems' legacy remains.
Bruce Buchanan was a philosopher and computer scientist at Stanford who collaborated with physician Edward Shortliffe to build MYCIN, the landmark medical expert system that could diagnose bacterial blood infections and recommend antibiotics. His specific contribution was developing the system's method for handling uncertainty — attaching numerical confidence scores called "certainty factors" to each rule, so the system could reason sensibly even when evidence was incomplete. A remarkable historical footnote: MYCIN was never actually deployed in hospitals, partly because of legal and ethical concerns about machine-made medical decisions, even though it had already been shown to outperform many human physicians in blind tests.
Intuition
Imagine you are a new employee at a law firm and you want to give advice as well as the senior partner, who has thirty years of experience. You cannot yet acquire that experience — but you can ask the partner to write down every rule they use: "If the client signed a contract without legal counsel AND the contract contained hidden arbitration clauses, THEN advise them to seek a settlement." Now, armed with that enormous written rulebook, you could theoretically give the same advice the partner would give — for every situation the rulebook covers. That is an expert system. The knowledge engineer is the new employee. The senior partner is the domain expert. The rulebook is the knowledge base. And the inference engine is the process of flipping through pages until you find a rule that matches your client's exact situation. The beauty — and the fatal weakness — of this approach is the same thing: the rulebook is only as good as what the partner remembered to write down. Ask about a situation the partner forgot to include, and the system stares at you blankly. The partner's deep, intuitive, "I just know it when I see it" judgment — the kind that cannot easily be put into words — was exactly what expert systems could never capture.
Historical Development
Origins · 1 minhow this was discovered
Why It Matters
Prerequisites
What This Leads To
See Also
MYCIN · Knowledge Representation · The AI Winter · History of Artificial Intelligence · Symbolic AI vs. Connectionism