There are some subtleties. Models like GPT2 are trained using many copyrighted documents whose authors could claim it is derivative of their work.