I think you're accurately describing a "regression to the mean" phenomenon, which LLMs have transmitted to areas involving programmatic control through text. I

I think you're accurately describing a "regression to the mean" phenomenon, which LLMs have transmitted to areas involving programmatic control through text. I recently used an LLM to help scaffold a piece of technical writing. I showed it to several of my peers (sharp and accomplished PhD students), and they immediately identified the paragraphs (even sentences in captions!) written by an LLM. I had spent a good amount of time prompt engineering Opus 4 to try and produce a coherent piece of content aligned with a narrative which I had already developed, but it was still _immediately obvious_ what was not consistent ... it stood out like a sore thumb to experts. By the the time we finished the content, there was not a sentence of LLM-generated writing in the final product. Possibly an incorrect extrapolation, but I think my experience here is the current status quo. Code is a bit easier for these systems, because the LLM can align against test suites, type systems, etc. But for writing ... you just can't trust the current capabilities to pass the sniff test by experts. After all, how does one commmunicate "coherence" as a test which the LLM might align its generations against? That's why you need a shared world model (with your audience) -- and if we could produce a computational representation of such a thing, we might have a chance at better coherent / consistent technical writing generations.