Raw LLM Responses

So first, you don't actually need to do that. If your AI has never seen a picture of Mickey Mouse tagged "mickey mouse", it will not generate a picture of that cartoon character in response to the prompt "mickey mouse". If it can, then they must have works that they knew were of Mickey, because they were labeled "mickey mouse". Their responsibility is to *not use* infringing works, not to *prevent infringing works from being created*. Second, our legal system makes a distinction of intent, and has the concept of fair use. A user who consumes copyrighted images specifically to train their model *not* to break the law is not intending to violate copyright. They are also operating within the confines of fair use, which is primarily concerned when derivative works compete with the creator's interest in the original works. Their actions explicitly work *against* competing with the copyright holders, so they should be just fine. So going back to the case of your crowd image: * The first question is whether they own the copyright to that image of the crowd, which originally belonged to the photographer. If not, they're violating the creator's copyright. * The hat in question would generally fall under "fair use". Pikachu is in the picture, but its incidental inclusion on a hat in a crowd shot does not compete with Nintendo's use of Pikachu. * If, however, the image is labeled "pikachu pokemon hat" in the training set, then it shows intent to use that image to generate derivative images of Pikachu, so it probably ceases to be fair use. * This sort of labeling would be necessary for the AI to generate another pikachu hat on demand. If it's un-labeled, the AI simply learns that pikachu-like clusters of pixels sometimes appear in crowds, but wouldn't know what they are. This is testable.