I have been sitting with a question about AI-generated semantic data models, and I don’t have a clean answer yet.
When you ask a large language model to produce an ontology or a domain data model, what are you actually getting? The model has already absorbed your domain from papers, standards, documentation and years of accumulated writing before you opened the prompt. So is the output new knowledge, or knowledge transcribed back to you in a format that looks like a deliverable?
When both sides are machines
It gets more interesting when the consumer of that model is also a language model, as it increasingly is. Then you have formalised a bridge between two systems that may already share the same understanding of the domain.
Which raises the question: what is the semantic model doing?
Descriptions and decisions
My tentative answer is that the models that still earn their keep encode something a language model cannot derive. Institutional decisions. Legal boundaries. The specific meaning an organisation has chosen for a term the rest of the world uses differently.
Not descriptions, but decisions.
A case for expressive ontologies
This is also why I think we swung too far in one direction. Ontologies used to be rich and demanding: classes defined through restrictions and axioms, reasoners that exposed inconsistencies and called out conceptual mistakes. They were dense enough to scare people away from the field.
Many modern models have gone to the other extreme: lists of concepts and properties, where even stating a domain or range feels restrictive. They are accessible, but they carry little of the logic a machine needs to act on context.
As systems become more agent-driven, the need for that logic grows. Agents need coherent, consistent context to reason with, and a list of terms does not provide it. Decisions, stated formally, do.
Is the formalism still the point, or has it become the receipt?
I am not sure yet where this leaves the discipline. But I am fairly sure the answer lies in modelling what only we can decide, and modelling it rigorously.