The reason I think it's a bad idea to differentiate between real or AI is similar to the arguments against "means testing" for distributing benefits. You don't want to put real victims in a situation where they're deprived of justice because they can't prove that their victimization was "real." Imagine a real CSAM criminal claiming a defense that they "thought it was AI generated." Do you want to give them that out?
If protecting those victims comes at a cost of punishing criminals possessing AI-generated CSAM with sentences equally as harsh as those for "real" CSAM, then it's a worthwhile cost to pay. They are still criminals, and they are definitely not innocent (unless they're being framed, but that's a risk with both real and AI images).
"I thought she was 18" doesn't work for physical sex either.
Saying "I thought this heroin was fake" is not a defense when caught with a bag of heroin, I don't see how this would be any different. It's not a magic out for anyone.
These are situational circumstances, and no prosecutor in the world would choose to prosecute those - but there is 0% chance you could get away with saying "oh I thought it was fake" if caught with CP on your phone.
It’s that you have a bag of fake heroin, you are then arrested for it because someone thinks that by you having fake heroin you are encouraging real heroin users to do more real heroin.