AI Recognition of Patient Race in Medical Imaging

This is a wild result, and I can't say I have any hypotheses for why this would make sense. It relates to this kind of theme that we discussed in the context of language models. What's baked in, hidden, not in sight, but hidden into these often otherwise inscrutable or under scrutinized models. So could you first maybe just tell us about why you started looking at it? And then tell us about what you found in the paper.