Why do you guarantee that? It seems obviously wrong, on a technical level.

The point I'm making is that even if we take it as a given that a shrewd enough AI could correctly distinguish sex at birth -- which I think is obviously impossible based on the appearances of many ciswomen and the nature of statistical prediction -- you'd still need a training data set.

If the dataset has any erroneous input, that corrupts its ability, and the whole point of this exercise is trying to find passing transwomen. Why would anyone expect that training set of hundreds of thousands of supposed cis women wouldn't have a few transwomen in it?