Thing is, LLMs level of compression of training set mean that effectively, under the same rules that say you cannot sell that database filled with copyright material, the LLM is fine to sell. Because you have to be able to meaningfully trace each claim to final output (weights). For example, for some older stable diffusion model, it was calculated that each individual work addition or removal resulted in about 1-2 bits of change, meaning the same rules would qualify it as not derivative work.

However, because it is an issue with (at least historical) goals of copyright law, the common pattern that is evolving is that AI is not granted copyright of any work it generates, making it a bit of poison pill for some of the egregious ideas of corporate abuse. Not sure if the weights will be considered copyrightable either.