Well, replacing novel creative functions with derivative creative functions. That's the big change I see here; similar to the difference between digitally editing an image vs. applying a stock sepia filter to it. Yes, we can use a model to regurgitate a mish mash of the data it was trained on, and that regurgitation might be novel in that nothing like that has been regurgitated before, but it will still be a regurgitation of pre-existing art. To some degree humans do this too, but the constraints are infinitely different.