The success path is the easy half
Designers spend their time on the moment the feature works. Users form their opinion in the moments it doesn’t — when it is slow, unsure, wrong, or unavailable — and those states are usually an afterthought or a spinner.
Show that it is working, specifically
Model responses take seconds, which is long enough for a user to assume something broke. Showing what is happening — reading, searching, drafting — buys patience that an undifferentiated spinner does not.
Make uncertainty visible without theatre
Where the system is unsure, saying so plainly beats a hedged answer that sounds confident. Giving the model a supported way to decline only works if the interface has somewhere to put that answer.
An interface with no way to say “I don’t know” forces the model to guess.
Always leave an exit
Every AI interaction needs a visible route to a person or a manual path. Hiding it to boost deflection produces the frustration spiral that escalation is supposed to pre-empt.
Make correcting it easy, and capture it
Editing rather than regenerating is faster for the user and produces a labelled example for you. Those corrections are the most valuable data the feature generates, and they only exist if the interface invites them.
Show the source
Displaying what the answer was based on lets a user verify it themselves, which converts an unverifiable claim into a checkable one and is the practical half of explainability.