There's no question about it. Llm output is not copyrightable. Which is moot anyways, because training on copyrighted data is completely legal.