You are misunderstanding the purpose of transformation as a Fair Use criteria; it's there to prevent using copyright claims to quash parody and critical (as in

You are misunderstanding the purpose of transformation as a Fair Use criteria; it's there to prevent using copyright claims to quash parody and critical (as in criticism) discussion of the copyrighted work, so for instance using copyright to attack a journalist who quotes shitty writing in a book review, or a comedian who alters a quote to make fun of a Disney character, etc. Not only is training ml models not a form of critical expression that enjoys 1A protection, but using the book itself to train a model isn't transformative; you didn't alter the book, you used the book verbatim to train on. That the model doesn't maintain the book doesn't matter; copyright doesn't only cover output (ie a derived work like said comedy skit), it also covers input (how the original work is obtained), and in this case they were not obtained legally (which means with a license for said use). > I’d argue that a corpus built for the training of an AI model is transformative since it doesn't detract commercially from the original works The reason OpenAI isn't arguing this is because they know that would be DOA. You cannot just package copyrighted works together or in a different way and claim that is transformative; there is actually direct case law about that because torrenters tried to claim exactly that, and lost. In fact, transformation is explicitly ruled in several cases to not apply to the entire original work being present.