I Asked ChatGPT to Build an Ontology. Here's What Happened.

Can ChatGPT actually make a good ontology?

In this video, I put an LLM to the test. I gave ChatGPT a Zillow listing for a residential property and asked it to generate a Turtle (TTL) representation of the data. Then I evaluated the resulting model—not just asking whether the TTL worked, but whether the underlying ontology was actually any good.

The result? The TTL was functional. ChatGPT could produce a usable RDF representation of the property. But when it came to the modeling itself, the results were much less impressive. The distinctions between concepts weren't always well thought out, and the ontology was underdeveloped, if not just making false claims.

My takeaway is that LLMs may be useful for generating a quick first draft of an ontology—especially as a starting point or a foil for human analysis. But if you're building a model that needs to be useful and maintainable over the long term, you probably shouldn't rely on an LLM to do the ontology engineering for you.

In this video, I walk through the generated model and evaluate where it works, where it falls short, and what this tells us about using LLMs for ontology development.

Can LLMs make ontologies? Yes.

Can they make good ontologies? That's a much more interesting question. (Probably not.)