[kelseyfrog]

> This study feels like it's saying "facial recognition AI model outputs vectors where female and male faces cluster closer to each other than not".

In this analogy, it would be like concluding that "androgyny does not exist."

I'm curious how they handled some of the complicated selection and coding issues around trans and intersex people. Were those groups excluded, or directed in responding in one way or the other? The study is large enough to expect a few of these cases.

The ability to perfectly separate the groups makes me suspicious of a data leakage or methodological error. Perfect accuracy rarely is.