Sufficiently advanced AI would be aware of the bias of its data, much more than any human ever could.
By the way, the IAT has very little credibility, and isn't based on the scientific method but thought up by sociologists. Ironic to use flawed human-made tools to call AI biased. The whole presentation is a joke: she uses a sample size of just a few Trump tweets, and then complains that the output would be biased (duh! sample size!); treated Tay AI tweets that were not written by Tay AI as proof of how "problematic" things can be (Tay could be made to repeat tweets made @ her verbatim, so many of these "problematic tweets" had nothing to do with its AI). She's also arguing that society needs to be reengineered by benevolent """scientists""" (social psychologists) and that people be bombarded with carefully-designed stereotypes to counter the stereotypes we've supposedly internalized. Disappointing to see this from a Princeton grad.
> Additionally, you still need to be really careful to write in ways that avoid furthering existing human bias.
Usually that involves not using emotionally-loaded "hand grenade words" that bypass peoples' rational minds, and using clickbait headlines written by malicious editors. AI could be made to avoid both.