[aftbit]

You still seem to be ignoring derived works a bit. For example, if I downloaded a copy of "Harry Potter" and replaced every "Harry" with "Harvey" and every "Potter" with "Plumber", my new book "Harvey Plumber" would be a derived work from the original, meaning I would not have the right to sell or distribute the new work, even though it wouldn't be exactly the same.

What if I did something similar, but rather than a simple Ctrl-H replacement, I asked an LLM to rewrite each paragraph in different words? What if I did the rewrite myself, by hand? Is there a difference? If so, why? If not, why not?

[dahart]

The right to make derived works is covered by copyright law. It’s not legal to play madlibs on something and call it a derived work unless you already have the copyrights. If the edit distance is too small, proving infringement isn’t hard. If you change more words, you’re still in violation of copyright on originality grounds, but it might be harder for someone to prove it.

The LLMs question is more complicated. If you ask AI for a rewrite of a specific work, that’s infringement on grounds of originality. It’s also infringement when you don’t have the rights to the work that you feed to the LLM. This is part of the debate over AI training, and is covered in the Copyright Office’s draft on generative AI under the Prima Facie Infringement section https://www.copyright.gov/ai/Copyright-and-Artificial-Intell...

AI companies aren’t arguing over derived works, they are trying to get approval to classify AI training as Fair Use, because they know they are infringing existing copyright law. The Copyright Office might end up allowing it and changing the law.