> considerabe case law on whether performing songs infringe on the author's copyright
But then you're no longer talking about "contains copyrighted material" you're talking about "actively reproduces copyrighted material in an immediately recognizable form".
> Generally I don't think copyright holders have agreed to have LLM ingest their works.
Aren't you essentially speculating about whether or not the training data was obtained in compliance with IP law?
The law should allow for training, but not reproduction. That would be compatible with the world that seems to be emerging.
You can train on Disney, but you can't produce Disney outputs.
Likewise, if you create some new IP using AI, the US will confer copyright so long as there was sufficient user input (eg. inpainting, editing, turning it into a long-form movie, etc.) Those new, sufficiently novel works should also be copyrightable.
Copyright should compel people to make good stuff and reward them when they do so. But it shouldn't hinder technological progress. There's a balance that can be struck.
But then you're no longer talking about "contains copyrighted material" you're talking about "actively reproduces copyrighted material in an immediately recognizable form".
> Generally I don't think copyright holders have agreed to have LLM ingest their works.
Aren't you essentially speculating about whether or not the training data was obtained in compliance with IP law?