Hallucination Testing
Hallucination testing checks whether an AI model generates confident-sounding output that's factually incorrect, unsupported by its source material, or entirely fabricated — like citing a paper that doesn't exist or stating a wrong number with total confidence. It's one of the highest-priority checks for any AI feature that presents information as fact.
Hallucinations are dangerous specifically because they don't look like errors — a model states a fabricated fact in the same confident tone as a correct one, with no built-in signal that anything's wrong. A traditional bug usually announces itself (an error message, a crash, a visibly broken layout); a hallucination looks exactly like a correct answer until someone checks.
Testing for it means deliberately probing the model's failure modes: asking about topics outside its training or knowledge cutoff, requesting citations and verifying they're real, asking questions with no correct answer to see if it fabricates one anyway rather than saying "I don't know," and testing retrieval-augmented generation (RAG) specifically to confirm the model's answer is actually grounded in the retrieved documents rather than drifting back to its own, possibly wrong, internal knowledge.
For QA teams, this usually means building a test set of known-answer questions, including some deliberately unanswerable ones, and checking not just accuracy but whether the model appropriately expresses uncertainty when it should.
