When AI sees a man, it thinks “official.” A woman? “Smile”

I'd also be interested how many of the women wear make-up intended to make themselves look younger vs men. Or How many of the (generally older) men are bald or have generic male haircuts versus the women. Or wear necklaces. Those might account for why "girl/kid" and "hair", "neck" might exist in the women's categories and not in the men's.

Since you can find out exactly why a label exists just by interrogating the NN - on that basis I'm disappointed by how shallow this article is, a whole load of speculation about "why", when it should be possible to find out. I feel like it's just playing on the readers biases/paranoia towards sexism in the tech industry, leading them to assume what the answer to these loaded questions are without Really trying to understand.

Another thing that annoys me is how often articles like this conflate bias in the data, and a poor choice of data by developers, with bias of the actual developers themselves. A NN has nothing to do with the phycology of it's creators, beyond choices of dataset, oversight etc. It essentially allows authors to make personal attacks against devs, as if a NN who says something racist was reading out a racist script written by a dev.