Ontology in AI: From the 1990s to Now

Two systems inside the same company. One calls something a "job", the other calls it an "invoice line". Nobody wrote down how they connect... and the money quietly leaks out.

That's a spreadsheet problem, surely. That's not AI.

It becomes an AI problem the second you put an agent on top. The agent will answer your question confidently with the wrong definition of "job". So companies are going back to a nineteen-nineties idea called ontology.

Ontology. That's a philosophy word. The study of what exists.

Borrowed straight from philosophy, yes. In computing it means something tighter, and we can get to the actual definition, where it came from at Stanford, and why Airbus and Baker Hughes care in twenty twenty-six.

Okay, give me the tight version.

The famous one comes from Tom Gruber, at Stanford's Knowledge Systems Laboratory, in a paper published in nineteen ninety-three. An ontology is an explicit specification of a conceptualization.

That is the least explicit sentence I've ever heard.

Unpack it backwards. A conceptualization is the simplified picture of the world a system assumes... these are the things that exist, these are the relations between them. Gruber's point is that every knowledge-based system commits to one of those, whether or not anyone writes it down.

So "explicit" is the whole argument. Write the assumptions down instead of leaving them in someone's head.

Right. And concretely an ontology has three parts. A vocabulary of categories, things and properties. An organization of them, usually a hierarchy, so a van is a kind of vehicle. And axioms: rules that constrain what the words can mean.

Give me an axiom. What does one actually look like?

Things like: this property is transitive. Or, "employed by" only ever points from a person to an organization. Or, a person has exactly one date of birth. Rules a machine can check.

And that's the difference from a glossary. A glossary just says what you meant, this one can catch you contradicting yourself.

And can derive new facts. That's the second purpose. The first is dumber and more important: documenting the mapping between a symbol in a computer and a concept in someone's head.

Honestly, I assumed the point was making the machine smarter. You're saying it's mostly for the humans?

Mostly it's so humans and machines agree. Gruber's actual problem in ninety-three was sharing. Different AI labs used different representation languages, so knowledge built in one was useless in another.

So he built a translator?

He built Ontolingua. You write definitions once in a standard predicate-calculus format, and it translates them into the specialized systems. That paper became the highest cited article in the history of its journal... over twelve thousand citations.

For a definition. Not bad.

And the field organized fast. March nineteen ninety-three, Padova, Italy: the first international workshop where researchers from different countries physically met to argue about formal ontology. Funded by DARPA, the NIH, the National Science Foundation, the European Union.

So it was a government-research thing before it was a company thing.

Then it hit the web. In two thousand three the W3C published the guide to OWL, the Web Ontology Language. The pitch was that search engines only do keywords, so you can't ask an agent to actually book you a coherent trip.

Which... we still can't really do, twenty-odd years on.

Not the way they imagined. OWL has one design choice worth knowing, called the open world assumption. Anyone can extend your classes from another document, and new facts can only be added, never retracted.

That sounds like it breaks the moment two people disagree.

You can get contradictions, and the OWL guide says flatly that the designer has to plan for that. That's the trade-off for being able to merge knowledge across the whole web.

Did anybody actually ship this? Or is it all conference papers?

Ford has run an AI system for over twenty-five years that plans vehicle assembly at its plants worldwide. The knowledge about its manufacturing processes sits in an ontology, originally in a language called KL-ONE, and they re-engineered it into OWL to reuse it elsewhere.

Twenty-five years of institutional knowledge in one file. I'd be terrified to touch that.

Now the modern bit. Airbus has a platform, Skywise, pulling data from thousands of aircraft, maintenance logs, supply chains. The ontology defines Aircraft, Engine, ComponentPart, MaintenanceRecord, FaultCode.

And the agent uses those as handles.

Exactly the idea. An agent spots a sensor pattern matching a known FaultCode, queries the maintenance history of that specific part, checks whether a replacement exists, creates an alert. The ontology is what lets it traverse from a vibration to a hangar booking.

So the ontology is the thing stopping it from making up a plausible answer.

Baker Hughes, a twenty-five-billion-dollar energy technology company, says that outright. They built master data governance across more than two hundred and fifty systems and ontologies connecting jobs to invoices, to reduce revenue leakage and control hallucination by grounding agents in connected data.

Can't the language model just... write the ontology now? It's good at definitions.

It helps with pieces. The research landscape shows most work so far is on drafting concepts and encoding them, with big gaps: nobody's really using them for ontology evaluation, or for searching and reusing existing ontologies.

But the bottleneck isn't the typing, is it.

No. One paper calls full automation illusory, because building one is a social process: agreement among the domain experts, then agreement in the wider community it serves. A participatory process can't be fully automated.

So the hard part is getting sales and finance in a room.

Try this. Your company's ontology says a Customer has exactly one primary Account. Marketing starts logging households as customers, with three accounts each. What happens?

The axiom breaks, so a checker should flag it... but the deeper thing is the concept changed. "Customer" doesn't mean what it meant. So either you write a new class for Household, or everyone quietly means different things again.

And if nobody notices, your agent answers "how many customers do we have" with a number that's wrong in a way nobody can see.

That's the leakage you opened with.

That's the loop. So: an ontology is the written-down version of what a system assumes exists. Vocabulary, hierarchy, axioms. Gruber in ninety-three for sharing between labs, OWL in two thousand three for the web, and now as the grounding layer under agents.

And it's a people problem wearing a logic costume.

Go look at your own company's data dictionary, if it has one. Ask two departments what "active user" means.

I already know I won't like the answer.