I think there might be actual technical reasons to (sometimes) be human-like. Saying something like "sorry, i meant ..." hopefully signals to the model that i m

I think there might be actual technical reasons to (sometimes) be human-like. Saying something like "sorry, i meant ..." hopefully signals to the model that i made a mistake, so it should not consider my previous input verbatim. When I say "thanks, that's good, and now ..." I assume it signals that whatever it just wrote is good and useful and it should base its next output on that (vs. whatever it had written previously).