also confabulation, made-up answer
A fluent, confident answer that is not supported by the model's input or by the facts. It comes from how models generate text, so systems reduce it with grounding, citations and a way to say "I do not know", not by asking the model to be careful.
You have done this if
The assistant cited a policy section that did not exist, and you added a check that every citation points to a passage that was actually retrieved.
Say it in a review
We treat unsupported claims as a measured failure: groundedness is scored per claim, and below the threshold the assistant declines.
On the AI Application map Model, Retrieval
Read Confidence is not evidence · Weak evidence should trigger recovery, not a confident answer