The interface has to communicate uncertainty
Conventional software is either right or broken. AI output is probabilistic, and an interface that presents every answer identically teaches users to trust everything equally — which means trusting the wrong things.
Showing uncertainty feels like admitting weakness. It is what makes the system safe to rely on.
Four things to always show
- What it is based on — the source document, the retrieved policy, the records used
- How sure it is, in terms a user understands rather than a raw score
- What a human still decides, stated plainly
- How to correct it, in one obvious action
Citations do more than any accuracy improvement
An answer with a link to the source is checkable in three seconds. An answer without one has to be believed or ignored, and users eventually choose ignored.
Make correcting easier than complaining
If fixing a wrong output takes more effort than working around it, people work around it and you never learn what was wrong.
- Edit in place rather than in a separate flow
- One-click “this is wrong” that captures context automatically
- No form asking why — ask later if at all
- Visible confirmation that the correction was received
Set expectations in the wording
“Suggested reply” sets a different expectation from “Reply”. “Draft, please check” is different again from “Generated”.
Small wording choices materially change how carefully people read output, and getting them right costs nothing.