A Nature study finds AI rewriting flattens linguistic diversity
Across seven datasets and more than 880,000 texts, LLM polishing cut the variance in writing complexity by 21 to 50 percent. The lost variation carries signal used in health, hiring and research.
A study in Nature Human Behaviour reports that having a large language model polish text reduces the variance in writing complexity by 21 to 50 percent, measured across seven datasets and more than 880,000 texts.
The finding is not that AI-edited writing is worse. It is that it is more similar — that passing text through a model pulls it towards a central style and removes the variation that distinguished one writer from another.
Why variation is not noise
The reason this matters beyond aesthetics is that individual variation in language carries information, and several fields use it.
Changes in a person's linguistic complexity over time are an early marker in cognitive decline research. Written expression is used in psychiatric assessment. In hiring, writing samples are read — rightly or wrongly — as evidence about a candidate. Across the social sciences, text produced by research participants is data, and its variability is the thing being measured.
If a substantial fraction of text has passed through a model before anyone sees it, the variance that these methods depend on has been partly removed before measurement. The signal is attenuated by an editing step that nobody records.
That is a measurement problem with no obvious fix, because the intervention is invisible. A researcher cannot tell from a text whether it was polished, and there is no reliable detector.
The homogenisation is structural
The mechanism is not mysterious. A model trained to produce fluent, well-formed prose has a central tendency, and asking it to improve a text moves that text towards it. Repeated across millions of documents, the aggregate effect is convergence.
The same dynamic appears elsewhere in the week's coverage. TechCrunch reported on the visual sameness of AI-generated restaurant menus — the same convergence in images rather than text, and more immediately visible because people can see it.
There is a compounding version of this that the study does not address but which follows from it. Models are trained on text from the web; an increasing share of that text has been model-polished; training on it narrows the distribution further. Whether that loop matters at current data volumes is unresolved, but the direction is not in dispute.
What the study cannot establish
The research measures complexity variance, which is a proxy for stylistic diversity rather than a full account of it. Two texts can have identical complexity scores and read completely differently.
It also does not establish harm. Reduced variance in writing complexity is a measurable change; whether it makes anything worse depends on what the writing is for. For a technical document that needs to be understood, convergence on clarity is an improvement. For a novel, it is not. For a research instrument, it is contamination.
The honest reading
The most defensible conclusion is narrow and still uncomfortable: any field that treats written language as data needs to account for the possibility that its data has been pre-processed by a model, and none of them currently can.
That includes the AI industry's own evaluations, which frequently use text produced by humans who had access to models. It also includes the growing set of enterprise tools built on analysing employee and customer writing — Conveo raised $50 million this month for AI-conducted customer research interviews in 15 languages.
The instrument and the contamination are increasingly the same technology.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters