That’s a fundamental misunderstanding of how diffusion models work. These models extract concepts and can effortlessly combine them to new images.

If it learns woman + crown = queen

and queen - woman + man = king

it is able to combine any such concept together

As Stability has noted. any model that has the concept of naked and the concept of child in it can be used like this. They tried to remove naked for Stable Diffusion 2 and nobody used it.

Nobody trained these models on CSAM and the problem is a dilemma in the same way a knife is a dilemma. We all know a malicious person can use a knife for murder, including of children Yet society has decided that knives sufficient other uses that we still allow their sale pretty much everywhere.

Getting an AI image generator to produce CSAM means it knows what to show

Not necessarily. Part of AI is blending different concepts. AI trained on images of regular children and nude adults in principle should be able to produce underage nudity. This is a side effect of the intelligence in the AI