Is that a bias? or do those images just look more like those categories at the pixel level?
There is a reason those words are the ones chosen by racists in the first place. Not just because they're hurtful, but because there is an objective similarity that makes the rest of the comparison seem, however shallowly, to validate their views.
Removing bias with censorship at the training set level is silly, and will likely hamper the AI's performance. Better to train the AI to not produce problematic output at the higher level, and ensure class membership in the training data is representative.