I'm not convinced it would distinguish the two. Perhaps the most concerning thing is that someone could create a nefarious image that's not obviously nefarious

I'm not convinced it would distinguish the two. Perhaps the most concerning thing is that someone could create a nefarious image that's not obviously nefarious to humans. I mean, we've all seen the very slight adjustments required to trick state-of-the-art image classification systems into thinking a cat is really a crocodile.