A compliance analyst used an AI assistant to draft an audit report. A reviewer discovers that the report cites a data-protection regulation section that does not exist, quoted in confident, authoritative language. No one tampered with the assistant or its inputs. Which risk does this represent, and what control BEST addresses it?

Hallucination is the tendency of generative models to produce fluent, confident output that is factually wrong or entirely fabricated, including nonexistent citations, statistics, and legal references. It arises from how models generate text by predicting plausible sequences rather than retrieving verified facts, so it can occur without any attacker. The appropriate control is human verification: a qualified reviewer checks every citation and factual claim against authoritative source documents before the output is relied upon, especially in compliance work. Bias is systematic unfairness toward particular groups and is not indicated by a single fabricated reference; retraining on balanced data would not prevent invented citations. Prompt injection requires malicious instructions inserted into the model's input, but the stem states no one tampered with the inputs, so filtering input addresses a threat that is absent. Data poisoning corrupts training data deliberately, and restoring a checkpoint assumes a known compromise that did not occur; even an untampered model hallucinates. Exam caveat: retrieval grounding can reduce hallucination, but it does not eliminate it, so human review remains the expected control for high-stakes outputs. Operational check: establish a policy requiring source verification for AI-assisted reports, record who verified each citation, and correct the draft before it is distributed.