I'd feel a lot better about that argument if we had sane copyright laws and anything older than 7-10 years was automatically in the public domain. Suddenly, netflix is looking a lot more valuable with just public domain works and there'd be a ton of public domain art to train AI models with. I suspect that the technology would still leave a lot of artists concerned in that situation though, because even once the issue of copyright is largely solved the fact remains that AI enables people who aren't artists to create art.

How would this compensation work? Let's say a portion of profits from LLMs that were trained on copyrighted work should be sent to the copyright holders.

How would we allocate which portion of the profits go to which creators? The only "fair" way here would be if we could trace how much a specific work influenced a specific output but this is currently impossible and will likely remain impossible for quite some time.

The big question is where "training an AI on this corpus of works and then either distributing the weights or performing the work via API" fall? Should the weights be considered derivative works? I personally don't think so and although the weights can be used to produce obviously infringing works I don't think this meets the bar of being a redistribution of the work via a funny lossy compression algo like some are claiming. But who knows? Copyright is more political than logical so I think the bend is really gonna be a balance of the tangible IRL harms artists can demonstrate vs. the desires of unrelated industries who wish to leverage this technology and are better for having all this data available.