So the training data would need to be manipulated to remove racial bias.

At which time the opportunity to add new biases would arise… nicely encoded in indecipherable AI networks.

> it does things. We’re just not completely sure how.