As opposed to illegal images, generating text does not physically harm someone. But it's known that some models will provide actual addresses of living people i

As opposed to illegal images, generating text does not physically harm someone. But it's known that some models will provide actual addresses of living people if a specific enough prompt is given. What if fragments of text from someone's sensitive blog posts start showing up between the lines of a dungeon crawl? I have to wonder about the legal implications of an AttnGAN or word-to-image type model trained on nothing but illegal pornography. Or one where the ratio of legal to illegal content the model was trained on is unknown and impossible to determine. There are some jurisdictions that consider synthetic images or text to be a victimless crime. At what point does the line become crossed and the images start becoming recognizable and realistic enough to be illegal at first glance under any usual circumstance, but are actually generated from nothing? How would people be able to prove the origin and veracity of such images for the purpose of submitting evidence of a crime?