The year was 1965. AI made its grand entrance in chemistry labs to find alien life on Mars. What?
I stumbled across this story while trying to make sense of a question I hear almost every week: 'Why aren't we using more AI?' It's usually asked with the quiet implication that the answer involves someone getting their act together - as if AI is sitting in a box somewhere, fully assembled, waiting for us to bother opening it. But the history of AI has a surprise in store. We've been here before. And the type of AI we built first - quiet, rule-based, invisible - is still running the world.
Back in the early 1960s NASA had a problem (one of many): it wanted to land an organic chemist on Mars. And it didn't know how to. Quite understandably, since it also did not yet know how to land an astronaut on the Moon. But why an organic chemist? To search for signs of life a prospective spacecraft landing on Mars would have to scoop up soil samples and analyse them for tell-tale organic molecules. The analytical method for this had to use an instrument called a mass spectrometer, a device which breaks down organic molecules into fragments. By measuring the mass of those fragments you can obtain a fingerprint of the molecules in question, allowing you to identify them. But interpreting the data (the fingerprint) coming back from such a spectrometer was tedious and needed a lot of expertise, which is why you needed an organic chemist to do the job. On Mars ... which was not going to happen. An unlikely trio of scientists took up the challenge to design an electronic brain which would be able to do the job ... by putting a chemist’s brain into a computer box.
This was not so surprising, as it was one of the hot research topics in AI at the time. The rage in those days was all about expert systems, and using logic and knowledge to codify the expertise of people into an artificial brain. Early AI research in the 1950s was obsessed with creating a "General Problem Solver" - a single algorithm that could use generic logic to solve any problem. It failed spectacularly because real-world problems require deep context and specialized knowledge. Enter computer scientist Edward Feigenbaum, geneticist Joshua Lederberg, and chemist Carl Djerassi. They set out to develop a computer program which could predict molecular structures based on mass spectrometry fingerprints, using a rule-based algorithm. It would avoid the General Problem Solver trap by focussing on a specific problem. The result of their collaboration was a computer program called DENDRAL, generally regarded as one of the first successful applications of AI in a real-world hands-on setting.
DENDRAL succeeded because it used a "Generate-and-Test" paradigm powered by Heuristic Rules. It mapped out an exhaustive list of all possible molecular structures for a given chemical formula. Instead of brute-force calculating every single possible structure, it applied "IF-THEN" rules "harvested" directly from expert human chemists. If a structural bond violated the laws of organic chemistry (the "rules"), the system pruned that branch instantly. The applied knowledge (coded in a computer language) allowed the number of branches which needed exploring, to be drastically reduced, thereby making the task manageable with the resources available.
DENDRAL proved that a computer did not need human consciousness to replicate human expertise; it just needed a vast, highly structured knowledge base and a set of carefully curated expert rules.
In the end, DENDRAL never made it to Mars. The onboard flight computers of the 1975 Viking landers were far too primitive to handle heavy, rule-based AI processing. Instead, NASA built the physical chemistry hardware Lederberg designed, flew it out to Mars, and beamed the raw digital data streams back across space. The AI code stayed behind on a Stanford mainframe, proving its worth not by finding aliens, but by automating the gruelling work of human chemists right here on Earth.
The success of DENDRAL paved the way for other systems in many business areas, and it triggered a real AI boom in the 80s and 90s. Go into any modern analytical chemistry lab, and the instruments - mass spectrometers, chromatography systems, and automated titrators - do exactly what DENDRAL did. They take raw data, run it against a computerized library of known chemical behaviours, apply carefully crafted business rules and output a validated compound profile. It is exactly the predictable nature of these operations which are the key to its success. When you make decisions on whether to release a batch of drugs to the market you don't want ambiguity or probabilistic responses, you need certainty and compliance. In a similar way, you want your ERP systems (think SAP or equivalent software platforms) to produce reliable financial statements or invoices, not hallucinated fictional customer orders!
These systems succeed precisely because compliance demands 100% predictability, not probabilities like LLMs.
And yet ... in today's AI conversations these systems are barely mentioned. Even stronger ... the popular conception of the history of AI holds that expert systems were a dead-end failure, and that LLMs are the only way forward. Why is that? There were several drawbacks to expert systems. The hardest challenge was to distil a consistent, accurate and complete set of rules from an expert's "knowledge". We humans often do not precisely know why we know something to be true or not. And as the scope and depth of these systems grew it became exponentially harder to keep the heuristic rules updated and avoid conflicting rules. These systems, despite their strengths, could not learn anything new on their own ... determinism can be both a blessing and a curse. And it is precisely here, in learning new knowledge, that LLMs really shine.
DENDRAL reminds us that deterministic, rule-based systems are not primitive software; they are the bedrock of operational excellence. The goal of technology in business isn't to remain flashy and experimental forever. The goal is to become so reliable, so repeatable, and so accurate that it blends into the background of daily operations. Expert systems didn't fail. They just won so decisively that we forgot they were AI.
That's worth sitting with for a moment. The next time someone in your organisation waxes lyrical about AI like it's a magic wand that appeared from nowhere - you now have a story to tell them. One that started in a chemistry lab in 1965, ended up on a Stanford mainframe instead of Mars, and quietly embedded itself into every LIMS, every ERP, every credit card fraud detection system quietly doing its job in the background right now.
The history of AI is messier, richer, and more interesting than the current conversation would have you believe. DENDRAL is just one chapter. In a future post I want to explore the flip side - where rule-based systems hit their limits, and why that opened the door to a very different kind of AI thinking. Stay tuned, and as always: if this sparked a thought or a question, I'd love to hear it.