Hallucination — the generation of plausible but factually incorrect content — is the central reliability problem of large language models deployed in clinical and regulatory contexts. A model that confidently cites a non-existent clinical trial, attributes an adverse event profile to the wrong drug, or describes a regulatory requirement from a jurisdiction it has confused with another is not making a trivial stylistic error — it is generating content that could mislead medical decisions or regulatory submissions. Ontological grounding addresses this problem at three distinct levels.
Hallucination at the Retrieval Level
Many hallucinations originate not in the generation step but in the retrieval step: the system retrieves documents that are superficially relevant but factually tangential, and the model generates content that is consistent with the retrieved context but inconsistent with ground truth. Ontology-driven retrieval reduces this problem by replacing fuzzy similarity-based retrieval with concept-exact retrieval: a query about compound X retrieves assertions about compound X specifically, not about structurally similar compounds that share some embedding proximity. The precision of ontological retrieval dramatically narrows the set of evidence that the model must synthesise, reducing the opportunity for cross-concept confusion.
Hallucination at the Generation Level
At the generation level, ontologies reduce hallucination by constraining the model's output vocabulary for factual claims. When the model is instructed to generate entity mentions using only concept identifiers from a specified ontology, it cannot invent drug names, disease names, or anatomical terms that do not exist in the ontology — the most common category of hallucinated biomedical entities. This constraint does not prevent the model from generating fluent prose; it simply anchors the factual nouns in that prose to a verified concept set.
Post-hoc Verification
The third level is post-hoc verification: each factual claim in the generated output is checked against the knowledge graph before the response is delivered. Claims that match knowledge graph assertions are marked as verified. Claims that contradict knowledge graph assertions are flagged for human review or automatically removed. Claims that cannot be verified in either direction are marked as unconfirmed. This three-category output — verified, contradicted, unconfirmed — transforms the AI response from a binary accept/reject object into a quality-graded evidence summary that human reviewers can act on efficiently.
