One of the things I ask AI to help edit my writing is a consistency pass, making sure my ideas hang together well. The real test is to see if other humans understand it.

I tell mine that line editing and formatting are the very last polish.

I think you make an interesting point.

I generally don't think LLMs writing prose for humans is a good idea (except for perhaps the most boring kinds of copywriting) but I feel a lot more positive about LLMs as reviewers, even if I don't necessarily use them myself often. My positivity comes from how relatively easy it generally is to validate feedback on a piece of writing you're intimately familiar with.

Gotta say, I don't relate to your concerns about "stray[ing] too far" _in the short term_. I think if reading LLM feedback has upset your original thinking that catastrophically you may have either been wrong (in which case, good that you received feedback early?) or writing with little intention.

However in the long term I can totally see how you could get into a habit of laziness, trusting the LLM to fix your stuff, or mindlessly applying fixes without really considering them.