This is not how ML works. You don't need real child abuse material in your training set to make a model capable of creating CSAM. All you need is depictions of any children and any sexual content.
Most publicly available sources of sexual content are adult, legal, don't have children
So if you feed Pornhub to your model, which most porn is actually well tagged, as is many ML data, you really have to go out of your way to get any children
Not saying is not possible, I am saying is easy to avoid. Not as an impossibility to separate both as you suggested above
If your training data have any absolutely innocent children pictures at all and also NSFW content with adults then model will be able to generate NSFW content with children. It's that simple.
Also even if you remove any children completely from training model will still be able to generate CSAM-like content because a lot of legit porn include jailbait models.
Most publicly available sources of sexual content are adult, legal, don't have children
So if you feed Pornhub to your model, which most porn is actually well tagged, as is many ML data, you really have to go out of your way to get any children
Not saying is not possible, I am saying is easy to avoid. Not as an impossibility to separate both as you suggested above