If the training data was a photo of every nurse in the world, then that’s what you’d expect, yeah. (And if was very low temperature/most average, you’d probably expect it to _always_ depict nurses as Asian). The training set isn’t a photo of every nurse in the world, though; it has a bias.