I think exactly the same things could be said for using LLMs to write text. Left to their own devices, the output appears functional, but is, in reality, clunky, inelegant, repetitive, and derivative, and the longer it is, the more narrative-breaking "bugs" you'll find, until the entire thing...