An AI trained on thousands of past decisions by compliance officers will learn their habits, including any inconsistencies or prejudices. Is it fair to build an AI this way, and what should be done to prevent it from repeating past mistakes on a large scale?