I guess proof of life videos mean nothing now.

but I sat my beloved husband down to look at them and he got all three in seconds. He's really good at noticing small details though.

On the second, I got two out of three, so improvement.

I agree with the heuristics given.

The so-called tells in this particular group are ears (AI has similar problems with ears and hands), moles, and facial hair. Another artifact I’ve noticed is that AI imagery tries to hide features it’s not good at rendering. Keep in mind that these images for comparison were selected by hoomans introducing another set of biases.

Is the real of the set the researcher is holding up disclosed anywhere in the text? I have my suspicions.

Whelp, it’s back to grinding through the Taylor-ish Swift music videos shown to the olds down at the home. Where people step off of moving trains and walk through trees.

see are the result of the AI user generating multiple image options - often dozens if not hundreds of AI images - and selecting the best. You don't know how many of the AI images produced to get to that good one were instantly recognizable as AI.

I think most AI image generators spit out four options at a time, so every single AI image online has a trail of unseen discarded images behind it. Every one of those discarded images a waste of electricity and water. I really can't think of good, ethical reasons to generate AI images. They're fake creativity using unethical tools trained illegally on stolen intellectual property.

Basically, almost everything you'll encounter online that's posted as an example of "how good AI is now" has been chosen just as selectively. There's almost always a long trail of complete or partial failures behind every AI success. And if you look at that trail, you won't see clear progression that would show learning by the AI model, as you'd usually see with a human artist making multiple attempts at an illustration. You'll just see fairly random options based on which part of the stolen dataset the machine latched onto.

The AI user might get better at choosing effective images. But that isn't creativity or art. That's shopping/selecting among the options.