UX in the age of AI: designing for probabilistic products
Traditional software is deterministic. Click a button, get the same result. AI-powered features aren't — the same prompt can produce different answers, and users notice. That gap is where trust is won or lost, and it's where UX has the most leverage.
Show the model its work
Confidence, sources, and reasoning steps aren't decoration — they're the affordances that let a user decide whether to accept, edit, or discard an output. Even a small “Show sources” toggle changes the tone of the whole product from oracle to collaborator.
Design for correction, not just generation
Great AI UX assumes the first output is a draft. That means visible edit affordances, easy undo, and side-by-side comparisons when a user regenerates. If regenerating feels risky, users stop exploring — and the whole point of the feature disappears.
A short checklist we use
- Is uncertainty visible when it matters (confidence, alternates, sources)?
- Can the user edit the output inline, without leaving the flow?
- Is there always a clear next action after the AI responds?
- Do we degrade gracefully when the model is slow, wrong, or unavailable?
AI features live or die on how the interface frames them. The model is the engine; UX is the steering wheel.