Ontology for Data Governance and Quality in AI Hallucination Propagation

Large language models generate false information, so-called hallucinations , without any internal truth-detection mechanism: they produce false claims with the same confidence as true ones. When downstream systems consume this output as valid input, they may execute actions that do not correspond to the real state of the world. This constitutes a critical data quality and governance failure, especially in critical systems such as healthcare and smart homes, where erroneous actions can cause physical harm, financial loss, or endanger vulnerable populations. The critical asymmetry is that the hallucination originates in the AI, but the erratic behaviour manifests in the consuming system. The AI does not know that it hallucinated; the system does not know that it consumed a hallucination. Existing ontology work on hallucination focuses on the model or on the engineering pipeline, but none formally models the full chain as a data governance problem: AI hallucinates, system consumes, system acts wrongly. This doctoral project proposes two complementary contributions: (i) a formal ontology, built with the OntoView method, that describes how AI hallucinations are consumed by downstream systems and how this translates into observable behavioural deviations, serving as a data quality and gover-nance instrument; and (ii) a verification method, based on the ontology, that specifies and evaluates boundary gates between AI output and physical execution, with direct impact on social well-being. The ontology and method will be validated by encoding reported cases, by expert evaluation, and by an empirical study on a simulated smart home testbed with scripted unsafe command scenarios.