AI features that show their working
Writing team:
Tomas Nyberg Engineering leadConfidence without evidence reads as a guess. Here is what users actually want to see instead.
Confidence without evidence reads as a guess
A model can be right most of the time and still not be trusted, because a bare answer gives a person nothing to check. The interface question is not how accurate the system is — it is what someone can do with the answer when they are unsure.
Show the inputs, not the internals
People do not want the weights. They want to know which of their own data the answer came from, and whether anything important was left out.
- Cite the source: the document, the row, the message it drew on.
- Say what was excluded, when exclusion changes the answer.
- Make the correction path as fast as the generation path.
Design for the wrong answer
Every AI feature ships with a failure rate. The ones that survive contact with users treat that rate as a design input: the recovery is built first, and the generation is fitted around it.
A test worth running
Give the feature to someone who knows the domain and watch them disagree with it. If disagreeing is harder than accepting, the feature will be accepted for the wrong reasons.
Ship the next release with a team that has done this before
We join in-flight products and leave them faster to change than we found them. Research, design and front-end, on one clock.
View services
