Could giving an AI an explicit map of concepts and relationships help it 'get' context better than plain text alone?

This explores whether giving AI an explicit map of concepts and how they relate, which is what an ontology is, helps it understand context. A direct caveat first: the collection has almost nothing on formal ontologies in the classic sense (taxonomies, OWL schemas, semantic-web vocabularies). What it does have is a cluster of work on a closely related idea. Structure that is imposed on what a model sees, or learns from, changes how well it handles context. Read that way, the corpus has some surprising things to say.

The strongest evidence comes from knowledge graphs, the closest cousin to ontologies here. One study turned paths through a medical knowledge graph into 24,000 reasoning exercises. A 32B model trained on them reached state-of-the-art results across 15 medical specialties, which suggests that composing structured relationships matters more than raw scale Can knowledge graphs teach models deep domain expertise?. The ontology isn't handed to the model at answer time. It's used as a curriculum, so the model learns the domain's concept relationships from the inside. A related line of work lets agents build their own concept graphs as they reason. These graphs settle into a state where about 12% of connections stay semantically surprising even though they are structurally linked, and that leftover surprise keeps producing new discoveries Why do reasoning systems keep discovering new connections?. So structure doesn't just constrain a model. Sometimes it's what makes new connections visible.

The most practical finding is about mapping what you *don't* know. Assistants become sycophantic and make things up partly because they have no representation of what remains unknown about the user. Adding a simple schema of labeled unknowns to the prompt cut harmful advice and sycophancy by 50–75% and roughly halved hallucination Do language models know what they don't know about users?. That is a tiny ontology of ignorance, and it does a lot. A similar instinct shows up in the ACE framework. It treats context as a structured, evolving playbook that gets updated item by item rather than rewritten wholesale, which stops hard-won details from eroding over many iterations Can context playbooks prevent knowledge loss during iteration?.

This matters because AI context is unusually slippery. Unlike a fixed software interface, it's a shifting mix of prompt, history, retrieved data and hidden state How does AI context differ from conventional software context?. It's assembled fresh for each request by product decisions, not by the model itself How do AI products decide what to remember between conversations?. Prompts also flatten the back-and-forth way humans build shared context into a single static frame How do prompts reshape the role of context in AI conversation?. Explicit structure is one way to put some stability back in. Even the design world is building its own ontologies here: Salesforce's glossary redefines terms like 'context,' 'memory,' and 'guardrails' as design surfaces for agent behavior Can familiar software terms gain new meaning for AI agents?.

There is also a sobering counterpoint. Farrell and colleagues argue that large models are lossy summarizers of accumulated human knowledge, closer to markets or bureaucracies than to minds that understand truth Are large AI models actually intelligent agents or social technologies?. On that view, an ontology doesn't give a model understanding. It gives it better-organized material to compress and recombine. The useful takeaway is that structure helps most where models are weakest, at the edges of what they know. Often the most valuable thing to map isn't the domain at all but the gaps.