I think it's deeper than that. I don't think copyright applies at all. > amount and substantiality of what is thereby made accessible to a public This starts

I think it's deeper than that. I don't think copyright applies at all. > amount and substantiality of what is thereby made accessible to a public This starts to get at it, but what really gets to me is the argument that it is the use of the data in training which copyright maximalists assert is wrong. Copyright doesn't protect style. If your model cannot be coerced into producing a duplicate of its original data, then there is no copyright question in the first place. Of course if it can, and if the amount and substantiality of the portions that it can reproduce are significant, then that protection goes away. But that's quite hard to do: there are cases where precise text reproduction can be done and has been demonstrated, but it's a real challenge to do it reliably and at length, except for some pathological special cases. To me the best analogy with what's going on today is the format shifting debate in the 2000s, except that to make that comparison you have to assert that an imperfect, unreliable destination format not intended for reproduction is equally harmful, and in the same way, as an mp3 rip of a CD. That's a hard argument to make. It's also very clear that the primary purpose of an LLM is exactly not duplication. You want them to be able to generalise over their input, not regurgitate it. They have several significant non-infringing uses. If they didn't, they wouldn't be worth the effort. Betamax would apply, in spades.