Defines hallucination as a confidently incorrect response, split into semantic illusion and confabulation by whether a semantically equivalent prompt recovers the correct answer, and kept separate from expressed uncertainty.
Overview
The word hallucination is used for many different model failures, which makes the results hard to compare. This paper argues for a stricter definition: a hallucination is a confident incorrect answer. A response that says it does not know is expressed uncertainty, not a hallucination.
Confident errors then split by a re-elicitation test. If some equivalent prompt recovers the correct answer, the error is a semantic illusion. If none does, it is a confabulation. Re-evaluating recent detection and mitigation methods under this split shows that they do not treat the three cases the same way.
Why it matters
Mixing these cases hides which failures a method actually catches. The paper recommends reporting semantic illusion, confabulation, and expressed uncertainty separately.