Yeah this tracks with what a lot of folks run into when they try to get an LLM to write longform essays without it drifting into “explainer mode.” They’re great at clarity but not so great at that messy human argumentation unless you feed it a lot of stylistic anchors.
I work at eesel AI (we’ve got a blog writer tool in our ChatGPT extension) and even there the trick is giving it enough voice cues or past content so it pulls from your actual rhythm instead of generic patterns. Otherwise everything starts to feel a bit too tidy and structured.
Going paragraph by paragraph helps but a small bit of grounding context up front usually makes the biggest difference.