Synthetic AI images (2) would have to be trained on a corpus of real images (1), except in the very unlikely event that many child-porn images are generated by

Synthetic AI images (2) would have to be trained on a corpus of real images (1), except in the very unlikely event that many child-porn images are generated by photo-realistic rendering of 3D animated models (CGI), where the 3D models are entirely created and animated by real content creators (i.e. entirely synthetic). So I think the default assumption for most realistic AI-generated images, is that they were generated from at least some real photographs. I am not an expert in AI images, but perhaps there is a way for AI to be trained on real pictures of children (innocuous, clothed, etc.), combined with real images of adult porn to create novel realistic child porn.