So I'm wondering if image-based NN generation will ever run across the same copyright issues as the source code generation of Copilot that's becoming a hot issu

So I'm wondering if image-based NN generation will ever run across the same copyright issues as the source code generation of Copilot that's becoming a hot issue at the moment. It's already known that Copilot was trained on GPL licensed code, and code can easily be searched to check for copying. What about a neural net that generates collages that resemble some obscure photographer's work that's under copyright? What about determining the copyright status of a model without knowing its source material? What will be the standard for determining copyright violations when, unlike code, it is not possible to coerce the model to output a near-1:1 duplicate of the original? And what about training a neural net on images that are illegal to possess in certain jurisdictions?