Traditional interfaces rest on a comforting assumption: the system is right. AI features break it — the system is usually right — and most of the UX work is designing for the gap. (METR even found users can't reliably sense AI's impact on their own output.)
What changes
- Outputs are suggestions, not facts. Signal confidence, invite correction, never present a guess as gospel.
- Latency is variable. Streaming and graceful waiting become core, not polish.
- Errors are different. The model doesn't crash — it's confidently wrong. Make noticing and undoing effortless.
Patterns that work
- 1.Human in the loop by default — draft, don't send; suggest, don't decide.
- 2.One-click correction — editing, regenerating, rejecting. Friction here kills trust.
- 3.Show your work — citations and sources build the trust probabilistic systems lack.
- 4.Design the empty and wrong states first — they're most of the experience.
Good AI UX doesn't hide that the system is uncertain. It makes that uncertainty safe, visible and easy to work with.
Sources
- METR — Developer productivity RCT
Written by ivector
Start a project →