Design systems for AI products: showing uncertainty honestly
Confidence, sources, and the moment a system should admit it does not know — interface patterns for products whose output is probabilistic.
Conventional interfaces present results as facts. AI products present guesses, and the interface has to carry that difference without drowning the user in caveats.
Show the source, not the score
"87% confident" means nothing to most users. A citation they can open means everything.
Make correction the fastest path
If fixing the output is harder than redoing the work, people redo the work. Editing should take one click and feed straight back into your labelling queue.
Design the refusal state first
Every AI feature needs a good-looking "I don't know". Designed last, it becomes an error toast; designed first, it becomes a trust signal.