What we learned building Murmur: how to show a model’s uncertainty without making people anxious.
Murmur, our listening experiment, taught us that the hardest interface problem in AI is not showing an answer. It is showing how sure you are.
Confidence is content
We tested five ways of displaying confidence. Percentages were precise but cold. Colour scales were intuitive but ambiguous. The winner was plain language with a quiet number beside it: "probably a blackbird".
Never hide the unsure
Low-confidence results are often filtered out. We kept them — clearly marked — because they were often the most interesting things the device heard.