Traditional software is deterministic. Press the button, get the result. Interface design has spent forty years refining the vocabulary for that world: affordances, states, confirmation, undo. None of it was built for a system that is usually correct.
Confidence is content
The most common failure in AI interface design is presenting a probabilistic output with the visual authority of a database read. The same typography, the same certainty, no indication that this one is a guess. Users calibrate to the interface, not to the documentation, and an interface that never expresses doubt teaches people to stop checking.
This does not mean decorating everything with percentages. It means the design has more than one register: a confident answer looks different from a draft, and a draft looks different from a question the system cannot answer.
Make the source reachable in one action
Trust is not built by claiming accuracy. It is built by making verification cheap. If every generated claim can be traced to its source passage in a single click, users develop an accurate mental model of when the system is reliable — which is far more useful than believing it always is.
Correction is a primary flow
In most AI products, fixing a wrong output is an afterthought — retype the prompt, try again, hope. But correction is the highest-signal interaction in the entire product. It tells you exactly where the system failed and what the right answer was.
- Let people edit the output in place rather than regenerate from scratch.
- Capture the correction as structured feedback, not a discarded diff.
- Show that the correction was received. Silent learning feels like being ignored.
Streaming changed the rules of latency
A spinner communicates one thing: wait. Streamed output communicates progress, direction and — crucially — gives the user the option to stop early because they can already see it is going the wrong way. Perceived performance is not a trick here; the ability to interrupt is genuinely valuable.
Design the moment the system is wrong as carefully as the moment it is right. Users will remember the first one longer.
The products that will feel trustworthy in five years are not the ones with the best models. They are the ones whose interfaces were honest about what the model is.