Generating arbitrary new images is extremely transformative, and reducing a zillion images to a few bytes each is pretty minimal. It is really fucking difficult to believe "draw Abbey Road as a Beeple piece" would get a commissioned human artist bankrupted, if they openly referenced that artist's entire catalog, but didn't exactly reproduce any portion of it.

For language models, it's even sillier. 'The network learned English by reading books!' Uh. Yeah. As opposed to what? If it's in the library, anyone can read it. That's what it's for.