Direct answer
An AI hallucination is a false or invented output presented as fact. An unsupported inference is a plausible conclusion that may follow from the source pattern but is not directly proven. Remove hallucinations; label inferences as candidates that need confidence, caveats or verification.
Why this matters
This distinction prevents a common over-correction: banning all inference. Good strategy needs inference, but public claims need clear proof boundaries.
How to use it
- Ask whether the statement is directly in the source.
- If not, ask whether it logically follows from the source pattern.
- Label the statement as inference, assumption or unsupported claim.
- Keep labelled inferences out of public copy until approved.
- Delete anything invented, false or source-contradicting.
Example
“Readers are asking for Google-style Q&A pages” may be source-derived from a project brief, search log or content tracker. “This will rank in AI Overviews” is an unsupported claim unless evidence and caveat are added.
Caveats and limits
Not every unsupported inference is bad. It becomes dangerous when it is presented with the same certainty as a verified fact.