Same but I focused on their expressions. Faces with expression “we are taking a picture of me” were fakes.

But that being said, all the pictures were insanely convincing and I picked fakes only because I knew I had to pick one and not because I knew one was fake.

+1 to the background trick. Faces were pretty convincing, but the generated backgrounds all seemed to have a generic filler appearance.

For me, the backgrounds were interesting for their _more salient_ features rather than their ambience, e.g. unexpected smears of color, textures that vaguely looked like real things at a glance (like fabric or nature) but wouldn't stand up to scrutiny. They reminded me of the typical "mistakes" that you see when playing around with image generators.

My heuristic in picking the fake one was to examine the fabric - the fake ones are all wearing what seems to be Dan Flashes.

For me its the eyes. A good number of AI generated photos I have seen have weird pupils or both eyes are not aligned to where they are looking.

I've used the lighting on the hair. Most of that time the specular highlight doesn't match the rest of the face.

In almost all cases, the generated picture had an uneven distribution of light reflecting from the eyes.

I also noticed the AI often got facial hair slightly wrong.

2 more potential things to check : digital artefacts on teeth and between hair and background Now they know what to improve ... Next round will be more difficult ...

And teeth. Definitely a few pictures where the AI gets the angle of certain teeth wrong from how they would look naturally.

Outside of any obvious anomalies, telling sign seems to be AI generates eyes symmetrical on the center of the screen.

yeah. but if this were randomized or corrected, it would be A LOT harder for me personally.

Skin is more glossy in theses IA photos than it should be.