Unless it can exclude the copyright outputs and provide something else instead of blocking the inputs. I'm sure there's AI that can check if a picture is close enough to something in their database of copyrighted characters built up from some kind of DMCA-like process of copyright holders submitting examples of their work to be blocked.

Indeed, but where does it stop? Looks like Potter? No go. Hmm, looks like an illustration of Pug? No go. Looks like Simon the sorcerer. No go. Hmm, looks like a wizard from Infocom's Sorcers get all the girls. No go.

The problem is that it regurgitates what already exists and if you really want to abide by all the permissions then there is nothing left.