i think the response would be that allowing racist tastes to persist in the successful model is itself an opinion embedded in your algorithm.

Its not that algorithm is an opinion, but that the goal achieved by it is defined with opinions (and thus, a successful algorithm is targetting those opinions)