I hope this is not the end of it. There is a lot of research that could be brought to bear on understanding this, and a lot of researchers that would be very in

I hope this is not the end of it. There is a lot of research that could be brought to bear on understanding this, and a lot of researchers that would be very interested. My guess is it boils down to something relatively trivial. 90% accuracy is not very good, it means there is some signal but there are also a lot of exceptions. The same way you can "predict" who is a criminal or gay or Democrat from a picture, there may be some loose stereotypes you can make that get you good enough correlation for an accuracy you can claim is good. But the model is not really learning a causal difference between the classes, it's more of a parlor trick. I'd love the opportunity to study how it works in this case.