As a thought experiment, let's say that the next version of stable diffusion is able to integrate large text datasets into the training set and can generate an accurate Mickey Mouse without ever having to be trained on an image of Mickey Mouse since it's integrated enough information from the text.
What then? Certainly an individual artist can't go and sell images of Mickey Mouse since it's still copyright infringement, but what claim would Disney have against the AI company?
I wrote in another comment that if you make the training of such models illegal regardless of distribution, it's essentially making certain mathematics illegal. That poses some very interesting questions around rights, whether others will do it anyways, and the practicality of enforcing such a rule in the first place.
What then? Certainly an individual artist can't go and sell images of Mickey Mouse since it's still copyright infringement, but what claim would Disney have against the AI company?
I wrote in another comment that if you make the training of such models illegal regardless of distribution, it's essentially making certain mathematics illegal. That poses some very interesting questions around rights, whether others will do it anyways, and the practicality of enforcing such a rule in the first place.