Making AI features feel trustworthy in the browser

Adding a model to a product is now the easy part. The hard part is the interface around it — because the moment software starts making suggestions, users need to know how much to trust them.

Review, not transcription

On ExamPro, the system extracts tags and keywords from examination questions. The temptation is to present that as finished work. We designed it as a review task instead: the tags arrive as proposals, the educator confirms or corrects them, and those corrections feed back into the engine.

The framing matters more than the accuracy figure. A user who understands they are reviewing will catch the mistakes. A user who believes the work is done will not look.

Show the work, not just the answer

Wherever a suggestion can be traced to something concrete — the phrase in the question that produced a tag, the record a summary came from — surface it. Users calibrate their trust from evidence, and a system that shows its reasoning gets more useful corrections back.

Present a model's output as a proposal and users check it. Present it as a result and they stop looking.

Latency is a design problem

Model calls are slow and variable in a way that ordinary API calls are not. A spinner that sits for eight seconds reads as broken. Streaming partial output, showing what stage the work is at, and keeping the rest of the interface usable while it runs are all front-end decisions, not infrastructure ones.

And the feature must degrade honestly. If the engine is unavailable, say so and let the user proceed manually. Silently returning nothing is how a promising feature gets switched off by the people it was built for.