I still haven't had time to look for the article, but your point triggered a question I can't answer. I'm told that the algorithms don't have the kind of memory that we use for writing, They predict the next word, but they don't plan the whole piece and they don't recall what's already been written when they get further into the piece. This means that a fully LLM-written novel will not and cannot have the kind of coherent throughline that a (competently) human-written novel would. (It also calls into question whether it's actually a novel, since a novel has narrative constraints that differentiate it from a book-shaped pile of pages. I have not done a lot of reading in narratology and I haven't read a fully LLM-written novel, so I can't say much more than that.)
Anyway, when I think through how I'd go about writing, say, an email when I've been told to keep it under 200 words, it's a multi-step process. I plan what I'm going to say, I say it, I check the word-count constraint, I check to see if I said what I'd intended, and I edit accordingly. I'm not sure current models do that kind of thing very well. It may be that it's a sufficiently recursive process that such a process would multiply time and energy demands to an extent that would make writing succinctly difficult.
