Even in situations where there's an objective reason for AI performing poorly with certain groups (like the worse light contrast on facial details with many pho

Even in situations where there's an objective reason for AI performing poorly with certain groups (like the worse light contrast on facial details with many photos of dark-skinned people), we go back to default and seek to cast moral judgment on someone for being racist, sexist, or something of the sort. Because you can't blame AI, we blame the people programming it, collecting data for it and training it. It seems we're trying to prevent a full-on moral panic that would be caused by the realization that "discrimination" is an objective need by the objective nature of the problem in some cases. For example there are about 14% black people in the US. If you have a FAIR SET OF DATA to train from, 14% of those faces would be black. No, they're not underrepresented, they're literally accurately represented. But this means the AI will be worse with those people. So what do we do? If we artificially up the "quota" and train with 50% black faces, the AI will now underperform with non-black faces. So then you need to train two networks: black faces and non-black faces, and then you need to "discriminate" between both, and pick the right network depending on the race you're working with. Math is racist sometimes.