I think in later cases we'll get some tests to apply about how much human intervention is required.
Who trained the LLM is probably not the issue, the courts would likely want to know about the training material. If I trained a model exclusively on Warhol art, and then had that model create new images in Warhol's style, I didn't do any of the creative work and probably don't get the copyright. Warhol's estate probably owns the copyright to the model generated images as they are derivative works.
I do think that a model trained on many different artists' works, with me providing substantial feedback to the model (and I can show the process), probably will at some point give me the copyright.
Somewhere there is a line:
- "Make a picture of a mouse." Probably not giving you copyright
- Using a model to erase a powerline in a photograph you took. Probably you own the copyright to the original image and the one without a powerline in it (regardless of how many other people's images the model was trained on).
- "Make a picture of a mouse, who is bipedal, wearing pink shorts, with a chip in his ear, wearing sunglasses, with scruffy whiskers, holding a surfboard, on his way to the beach to hit some waves." then updating with "make him shorter, give him blue sneakers" and then updating with numerous other tweaks until you get it just the way you want. Who knows where this lands?
I think that in the short term the courts are going to land on the side of "anything made by a model trained on existing artwork is derivative of the training set so you can't own the copyright, no matter how much you tweak it." I think eventually the courts will recognize there is some amount of input that makes the computer image the realization of a vision in your head, and not a derivative of the training set. Just how every individual musical note has been played before, but at some point, you put them together in an arrangement that is original.