You'd have to prove that the A.I. was discriminating based on a "protected class" and not on some other basis. But you have no insight into the A.I. or its training data. Nor do you have a comparable A.I. of your own to run A/B experiments that can prove discrimination. Now what?
