It's important to note that to produce an AI-generated image of somebody, they do not need to exist in the training data beforehand. You can use an image as a prompt, and an AI can create new works based on that prompt, without ever referencing the person from training materials.

Most of these CSAM files are not readily available on the clearnet, and I don't think even Elon is stupid enough to let an AI scraper run free on Tor. One would have to go significantly out of their way to locate clearnet sources of CSAM to include into the training data.

Some would say it's a distinction without difference, but I think it's important to understand how these AI images are actually created, if you want to adequately legislate them.