The problem with erasing biases is that you cannot look at any statistics. The internet and training set can be free of any form of -ism, and the models would still be expected to have biases. In fact it's something desirable, because statistical inferences are a valuable tool.

The AI won't care if some people get upset because it consistently recommends you get Mexican food instead of Italian when you're visiting south Texas. The weak link is humans not recognizing that that doesn't mean there cannot be good Italian food in south Texas. A logical hurdle I don't see AI having any problem with.