The constraint that shaped everything
A doctor has ninety seconds. If the interface asks them to verify the model's reasoning, they will stop using it. If it asks them to trust it blindly, they should stop using it.
Showing uncertainty without noise
We stopped showing confidence percentages. They were read as precision the model did not have. What worked instead: showing what the answer was based on, ranked, and letting the source do the convincing.
What carried over to other work
Every AI feature I have designed since starts with the same question: what does this look like when it is wrong? If the answer is 'the same, but wrong', the design is not finished.
Useful links from this article