remember Watson in jeopardy? Even when it got the question wrong you’d often see the right answer in the list of possibilities it showed.

Users need a way to train/override the ML to “fix” mistakes.