I think his point is that if people set the objectives, the AI will learn what people like, and that is going to be discriminatory in some way.
The difference is that a person with some flaw can compensate by having other qualities and still be popular. But if there is a hard rule, then that's not possible.