I’ve done a lot of stuff with generative models called GANs, like StyleGAN (2?) which I believe these pictures to be from. My main focus is the “hair bubble effect”. This works best for people with longer hair, which is why I had 30% wrong in this test. Basically, by starting at images generated by these models for a long time, I started noticing that it is bad at creating the few loose hairs that stand out from the main pack. These plucks of hair often seem to go around some invisible “bubble” or weirdly flow together with the background. So my main point of focus is often the transition between hair and background, or just the hair in general, since that’s where it’s most likely to mess up. But the images picked here were also intentionally picked to be the most confusing according to the rest of the article, so it’s not that weird that these are hard to classify. Some of the real ones looked extremely AI to me, and it was only after the first false positive that I got a lot more careful with labeling some as “AI” than I normally would.

Example, the strands of hair here (though admittedly the effect is not very convincing here):