Can AI Generate an Ontology?

Yes, mostly, for a first draft. Point a current model at a schema or a corpus and it will generate an ontology in minutes, with a sensible class hierarchy and plausible relationships. If you have tried it recently you already know this. A page arguing otherwise would be arguing with your own experience.

These products ship today. Vendors describe building "a working model of your entire business, not just a catalog of your tables… automatically from the signals your business already produces". They promise "no consulting engagement or manual build to wait on", going "from prompt to deployed model in hours, replacing the six-to-thirty-six-month projects".

The harder version is worth stating too, because a rebuttal that ducks it is worthless. At least one of these pairs generation with validation. One product page claims "251 enforceable rules, two architect reviews, and an agentic loop that proves the model before it ships." A cloud provider's prescriptive guidance agrees that "LLMs have made it easier than ever to facilitate the creation and management of ontologies and associated rules/axioms… tasks that previously required months of manual knowledge engineering."

So the question is not whether generation works. It is what you have when it finishes.

What the measurements actually show

The best published measurement available compared LLM-generated ontologies against a human-made reference, across three models and with a small qualitative study alongside it. Property F1 ranged from 0.12 to 0.27 across every configuration tested. Classes fared better, and far less consistently: the best run came out several times the worst. The qualitative verdict was mixed. The output formed "a strong and well-structured foundation" but came out "shallow, flat", with "too many top-level classes and insufficient depth in the hierarchy." (A Comprehensive Benchmark for Evaluating LLM-Generated Ontologies, CEUR-WS Vol-3953.)

Properties are the part that matters here, and the gap is not subtle. A class hierarchy is a vocabulary. The properties are where the meaning lives: what relates to what, under which constraints. A model that names your concepts well and connects them poorly has produced a glossary with ambitions.

Two caveats travel with that benchmark and both belong in the open. It used 2024-generation models. So it is the last published independent measurement of its kind, and it says nothing about what a frontier model does today. A rerun would settle that. Second, four of its five authors work for a vendor in this space, with the fifth at a research institute. The work is peer-reviewed and the figures are the figures. A reader deciding how much weight to give it should still know who ran it.

Read the two numbers together, because the gap between them is the finding. Models get the nouns roughly right and the relationships badly wrong. A flat list of classes is a vocabulary. In an enterprise, an ontology is the agreed answer to two questions. What are the organization's core things, and how do they relate? Every person, system and agent then resolves those questions the same way. The full definition is here. The relationships and the constraints are no decoration on top of the class list. They are the part that does the work. They are also the part the measurement says is weakest.

If you suspect current models have closed that property gap, you may well be right. Nobody has published a rerun, and a rerun is what would settle it. Until somebody does, this is the best number anyone has, and its age is a real weakness.

Why is generated structure inconsistent between runs?

Output can be good and still not be the same twice. A practitioner states it plainly: "Run it again on the same text and you will get different names for everything. If the terms used are unpredictable, then your KG is of little use to anyone else."

Academic work finds the same from another direction: models "refer to entities in slightly different ways, even when presented with the same text, prompt, and instructions." Given "Alice met Bob in New York City," different models returned "New York," "New York City," and "NYC." Audits of generated graphs find the predictable consequence: a node count well above the number of things described. That is why building an ontology an AI system can use treats entity resolution as its own step.

This is not a quality problem a better prompt fixes. It is a determinism problem, and determinism is what makes a model something a second system can rely on.

What generation structurally cannot produce

The broad version of this claim is false, so here is the narrow one. A generated model is a description of what is already happening. It is not a record of what anyone agreed.

It cannot tell you who decided a definition, what that definition replaced, which systems are still pinned to the old one, or when it expires. Continuous automatic refresh is no governance mechanism. It is a system designed to change a definition without anyone being asked. That is exactly right for a search index. It is exactly wrong for a number that goes to a regulator, a board or a patient.

This is a property of the substrate, and there is evidence for it. Major vector-store products provide no native way to verify where a stored embedding came from. A 2026 paper on embedding-store provenance documents the gap and proposes a cryptographic defence for it.

The sharpest argument against this position comes from a data catalog vendor: "What matters is not the formalism of the representation, but whether the context is accurate and whether it self-corrects when the business changes." It deserves a direct answer. Self-correction is a property you want in a description and cannot have in a commitment. A model that quietly corrects itself is ideal for retrieval. It is unusable as the definition behind a reported number. The whole value of that definition is that it did not change while nobody was looking.

Are these two separate objections?

They are, and they are routinely collapsed into each other.

The table makes two things obvious that are worth saying in words. A vendor can ship a careful review trail and still own the runtime your model lives in. A vendor can ship open standards and still refresh a definition without asking anyone. Both of those are happening in the market right now, and neither argument covers the other. TopQuadrant's claim is to the bottom-right cell, which is a more testable statement than either objection alone.

It is also worth being accurate about how the automation vendors behave, because the lazy version of this argument is a misquote. Every one of them hedges with human review. One states that extraction "does not publish anything on its own… Nothing goes into production because a model proposed it." Another requires a "proposal-first model so users can review them before applying them", by design. One platform vendor argues against LLM-invented semantics outright. Its position is that "the shift from BI to AI agents requires a semantic foundation grounded in data usage, not LLM assumptions". That is a different argument again, and it should not be lumped in with the others.

And the generation camp reaches the authority conclusion on its own. In September 2026 a lakehouse vendor published that you do not need a formal ontology built from scratch. Seven days later it published that some concepts are too important to leave to inference, and should be defined once by your own experts. The second position is reasonable. The interesting part is that it arrived unprompted.

So what is the human step?

Narrow, and nameable.

Review the properties. That is where the measurement says the output is weakest, and it is the cheapest place to spend an expert's hour.

Resolve the entities. Decide that Apple Inc and AAPL are one thing, and record that decision so the next run does not re-derive it.

Record the agreement with an owner and a date. Without those you have a description nobody is answerable for.

That last step is the one that converts output into an asset, and it is worth being precise about why. A generated model is a reading of the evidence. It becomes ground truth at the moment someone with the standing to decide says so. The record of that decision is the review trail: who proposed it, who approved it, and what it replaced. AI can propose definitions, relationships and rules all day. None of it is ground truth until the people accountable for it agree.

What to keep from the generated draft

• The structure. Class hierarchies come out usable, and rebuilding them by hand buys nothing.
• The candidate relationships, as proposals to be confirmed.
• The diff. Keep the generated version in source control so the difference between what was proposed and what was ratified stays visible. That diff is the most valuable artifact the exercise produces.
• Nothing unreviewed. A property nobody looked at is a property nobody agreed.

Express the reviewed result in open standards. A generator's native format dies with the generator. RDF, OWL, SKOS and SHACL are readable by the next system, which matters more than usual here, because generation tooling in this space turns over fast. Reuse also cuts the problem down before generation starts. Extending one of TopQuadrant's 300+ reusable standards-based models gives the generator less to invent and the reviewer less to check. Keeping the result accurate afterwards is ontology management, run on a platform built for the review loop.

Take the generation. It beats people at first passes and at change detection, and refusing it on principle is just slower. The method is in how to build an ontology for AI and how to create one from an existing database. Whether you need one at all is a different question.