Hallucination

A hallucination is a generated claim that is not supported by the evidence available to the system.

Fluent language is not evidence. A model can confidently produce an incorrect date, source, API, or customer policy.

Controls

• retrieve authoritative evidence at request time;
• ask for abstention when evidence is insufficient;
• require citations for consequential claims;
• validate structured outputs and tool arguments;
• evaluate groundedness and correct abstention on representative cases;
• keep humans in the loop for high-impact actions.

These controls reduce risk but do not create a guarantee. The acceptable residual risk depends on the workflow.