I can write a book about a boy wizard's adventures wizard school and that's legal, but if I call them Harry Potter it isn't.
I can create Harry Potter fanart and distribute it online pretty freely - but slap it on a mug and sell it, and that's illegal.
I can record an audio description of a painting that's as detailed as I like and it's legal to distribute - but take a photograph of the same painting and it's a derivative work, no matter how artistic my choice of camera settings.
We don't really have any prior examples that are precisely like these huge ML models trained on copyrighted data - and depending on which imprecise analogy you choose, you can come to a different conclusion.