The ultimate issue is that the models don't encode the training data in any way that we historically have considered infringement of copyright. This is true for both transformer architectures (gpt) and diffusion ones (most image generators). From a lay perspective, it's probably good and relatively accurate for our purposes to imagine the models themselves as enormous nets that learn vague, muddled, impressions of multiple portions of multiple pieces of the training data at arbitrary locations within the net. Now, this may still have IP implications for the outputs and here music copyright is pretty instructive, albeit very case-by-case. If a piece is too "inspired" by a particular previous work, even if it is not explicit copying it may still be regarded as infringement of copyright. But, like I said, this is very case specific and precedent cuts both ways on it.