Considering that every image generation model out there tries to censor your prompts/outputs despite trying their best not to train on CSAM... you don't need to train on CSAM for the model to be capable of generating CSAM.

Not saying the models don't get trained on CSAM. But I don't think it's a foregone conclusion that AI models capable of generating CSAM necessarily victimize anyone.

It would be nice if someone could research this, but the current climate makes it impossible.