> Its obviously impractical for a number of reasons, I just find it frustrating because based on the last century of IP law that should be what's happening.

Eh. Most LLM output is significantly transformative. The challenge here is that some of the output is similar enough to a work in the training data that it could reasonably be considered derivative. And evaluating whether it is when you have both of them in front of you isn't even the hard part.

The real problem, from the perspective of LLM users, is that you don't know when it's happening because you haven't seen the original work to notice how similar it is to a particular output, and now maybe you have problems if you start making copies of it.

> The big corporations won the fight on file-sharing so they ought to at least be accountable to their own legal arguments.

Those big corporations are the likes of Sony and Comcast. They're the ones who would prefer to be able to sue OpenAI. But that's also the thing which is trash, because if the money went only to them then it's just propping up the incumbents while spitting on small independent content creators who get nothing. Whereas if you actually included everybody then we're back to ~$3/year and it's not even worth the administrative costs compared to just having them pay ordinary taxes on their profits to the extent that they eventually have any.