Designing for the answer that might be wrong | Brain Quest
{{ post.deck }}
{{ sec.h }}
{{ para }}
- {{ li }}
{{ sec.quote }}
The market need
Most AI features fail in the same place: not at the model, but at the moment a user has to decide whether to believe it. Trust is the conversion step, and it is almost always designed last, by whoever is left.
The commercial consequence is measurable. Features that surface uncertainty well get used repeatedly for real work. Features that present every output with the same flat confidence get tried once, caught being wrong, and quietly abandoned.
Three controls, in order of importance
After a few products, the pattern is consistent. Everything else is decoration on top of these.
Undo. Nothing the model does should be irreversible without a confirmation the user actually reads.
Sources. Not a citation count — the specific sentence the output leaned on, one click away.
Confidence, expressed as behaviour rather than a percentage: hedged language, a visible gap, a request for clarification.
Percentages are a bad way to show uncertainty
We shipped a confidence score once. Users read seventy-eight percent as “correct” and forty percent as “broken”, and nothing in between changed a single decision. The number gave the interface an air of precision it had not earned.
What worked better was changing the shape of the output. A confident answer states the thing. An unsure one asks a question first, or shows two options side by side, or fills in the field but leaves it visibly unconfirmed. Users read behaviour fluently. They do not read probabilities.
Do not tell someone the answer is seventy percent likely. Show them the two answers and let them pick.
Latency is part of the trust surface
A spinner is a promise of a single correct answer. A stream is a demonstration of work in progress, and users forgive a stream that corrects itself in a way they never forgive a spinner that resolves to something wrong.
Where the work genuinely takes time, narrate it. “Reading twelve documents” is not a progress bar; it is an explanation of what the answer will be based on, delivered while the person is already waiting.
The failure state deserves a design
Every AI feature will be wrong in front of a customer. The question the design has to answer is what that person does in the next fifteen seconds. If the honest answer is “start over”, the feature is not finished — there should be a path from the wrong output to a better one that does not discard the context they already gave.