Yeah, this is a valid point, if this is the exact case or not I don't know, but a lot of people don't realize a lot of the weird biases that can appear in the training data.

Like that AI trained to detect ig a mole was cancer or not. A lot of the training data that was cancer had rulers in them. So the AI learned rulers are cancerous.

I could easily see something stupid like angle the picture was taken from being something the AI erroniously assumed to be useful for determining biological sex in this case.

It's possible to manipulate an image in a way that the original and the new one are indistinguishable to the human eye, but the AI model gives completely different results.

Like this helpful graphic I found

Or... edit the HTML...