Field note · 2026-09-24
The paper that sounded like a model talking to itself
My paper drafts were accurate and read wrong. They sounded like a model working through a problem, where a paper is a person explaining a result to another person.
- 139
- runs of bold type across the drafts2026-09-24
- 9
- figures listed in the drafts, none of them drawn2026-09-24
- 26 + 33
- style changes in the two papers, each logged before and after2026-09-24
- 0
- numbers that changed during the rewrite, checked by script2026-09-24
What I saw
Three paper drafts came back from review and none of them sounded like me. Every section closed on a moral. Bold type was doing the arguing: 139 runs of it. The phrase "rather than" appeared about forty times. Nine figures were listed and none had been drawn. The text also used words my own project had invented, so a number was a "headline" and a result was "anchored," and a reader from outside would have needed a glossary.
What caused it
A model writes the way it reasons, and it leaves its working in the text. I ended up with three names for what I was seeing. Thinking out loud: the draft narrates the path to the result. Private vocabulary: nicknames that grew up inside the project leak into prose meant for outsiders. Typography as rhetoric: bold and short punchy sentences stand in for evidence. There is a fourth thing underneath. Models don't think in figures, so they describe in sentences what a scientist would show in a plot.
What I changed
I built the style guide from my own published writing. My 2020 paper is plain, uses the passive for procedure, and leads with figures: ten figures and four tables in about eight thousand words. That became the voice the agents write to, with a before-and-after table for each failure and a checker that runs on every draft.
Then the papers were revised against it. One took 26 style changes and 4 changes to claims; the other took 33 and 2. Every passage is logged before and after with the reason. A script compared every number in the text with the typeset version afterwards, across 1,090 and 1,656 tokens, and found no discrepancies. The rewrite changed how the papers sound and left what they say alone.
I also settled disclosure. The papers say in one sentence, in the acknowledgements, that language models helped draft the text.
What I'd tell another team
People in my field don't object to a model drafting a paper. They object to how it sounds, and they are right to, because prose that reads as unattended makes a reviewer wonder what else was unattended. Give the model your own published writing as the voice, and ask for figures first. Translate every internal nickname before it leaves the building. Then read the draft aloud to a colleague and cut any sentence that is there to sound good.