I wonder if racists get friend suggestions for people who aren't their "race" in appearance terms; I'm imagining that could be an emergent selection not explicitly encoded for. If that were the case ... would it be so different to people making that selection. It's a question of allowance vs. instigation I suppose ... but if you knew that behaviour was likely to emerge then the line gets blurred somewhat ...

Rarely. I operate several facebook profiles that exclusively look at very racist or politically extreme people, and guess who FB keeps recommending as potential friends.

Why?

I study people like the other person who replied to you (whose comment has since been removed).

How are algorithms a "layer of abstraction" for guidelines given to human moderators?

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.

People need that hope.