DENDRAL was a research program begun at Stanford University around 1965, led by the computer scientist Edward Feigenbaum, the geneticist Joshua Lederberg, and later the philosopher-turned-computer-scientist Bruce Buchanan. Its goal was narrow and concrete: take the readings from a mass spectrometer, an instrument that breaks a molecule into fragments, and work out which organic compound produced them. In doing so it became the program that founded a whole paradigm. Where the earlier Logic Theorist had searched for proofs in pure logic, DENDRAL was the first system to capture the specialized know-how of human experts in a real scientific field and put it to practical work, and it is widely called the first expert system.
The key idea was to separate general reasoning from domain expertise. DENDRAL could in principle enumerate every chemical structure consistent with a molecular formula, but that space was far too large to be useful. The breakthrough was loading the program with the heuristics, the rules of thumb, that experienced chemists use to rule out impossible or implausible structures quickly. This made it the first clear demonstration of what Feigenbaum would later call knowledge engineering: that the power of an AI system comes less from clever search and more from the quantity and quality of expert knowledge built into it.
The Stanford report “On Generality and Problem Solving: A Case Study Using the DENDRAL Program,” issued as Stanford memo AIM-131 and reprinted in the Machine Intelligence series, lays out the project’s own reflection on this tension between generality and expertise. The authors found that a program broad enough to handle anything performed poorly, while one packed with chemistry-specific rules performed at the level of a specialist but only within its narrow domain. That trade-off would define expert systems for the next two decades.
Why business readers should care: DENDRAL set the template that MYCIN, XCON, Cyc, and the entire 1980s expert-systems industry followed. It showed that AI could deliver real commercial and scientific value by encoding human expertise as rules, and it also exposed the recurring weakness of that approach, namely that such systems are expensive to build, hard to maintain, and brittle the moment a problem strays outside the knowledge they were given.