That's not necessarily true. You could have all sorts of training models that didn't rely on any real children at all. You're just assuming it was trained on actual child porn. Other ways it could be trained:

• Trained on images of adults that look younger than 18. You could commission a bunch of pornography by people with whatever medical condition this guy has and use that as your training data.

• Trained on pornographic images of adults and regular images of kids. (This one is questionable legally, even though again, no kids were harmed in the making of it.)

• Trained on entirely on images of adults, reviewers are tasked with trying to identify the age of the person in the image generated, and using that as training data.

If someone was deliberately trying to create a legal child porn engine (what a horrible phrase to type), they would be sure to do it without using any actual child porn as the training data. There would be little point otherwise.