Why does it matter than some tiny fraction of AI training data is child porn? Does it taint the whole model, or just when it generates child porn, and why? Does the fact that it has trained on other problematic content, such as photographs of real murders, mean that generated pictures of people being killed is more problematic than other ways of making those pictures? (Not to the point of illegality, since possessing pictures of a murder is not usually illegal). "Respect for the victims" is a vague thing that could be used to justify more or less anything; it's not a sufficient principle, in my opinion, for curtailing freedoms. Not without something more concrete and objective.

CP is a synonym for CSAM, although CSAM is more respectful. This is just semantics.

"Semantics" means "what stuff means", and what CSAM means specifically is Child Sexual Abuse Material. It's not just a synonym.

I’m saying that if it’s something ONLY used for that, then its harm is undeniable.

You're going through a number of situations which are harmful, and I don't disagree about that. But that doesn't prove that every situation is harmful. Just having child porn on your computer doesn't harm anyone. Hell, jerking off to it does not - as far as I am aware from the present best research - harm anyone.

And what would stop a predator from saying that a footage or picture of CSAM was AI generated and get away with it?

If you can't prove harm, then it shouldn't be illegal, I don't think. If there's proof of harm, then it doesn't matter what the predator says.