They’re really, really bad at context. The main failure case isn’t making things up, it’s having text or image in part of the result not work right with text or image in another part because they can’t even manage context across their own replies.
See images with three hands, where bow strings mysteriously vanish etc.