The population effect of convenient editing
The July 29 preprint “Linguistic Monoculture in LLM-Assisted Language Use” models how writing diversity changes when many people repeatedly draft and revise with the same language model. Authors and models are represented as distributions over linguistic features under three settings: a fixed shared model, a shared model retrained on author output, and personalized models.
The predicted long-run outcomes
In the framework, a fixed shared model pulls authors toward one norm. Recursive updates can move that shared norm without restoring pairwise spread when conformity is common. Personalized feedback can preserve multiple author-model equilibria and nonzero diversity.
Clarity, legibility, and fluency can make conformity individually rational. The model argues that people may nevertheless conform more than is socially optimal because they do not internalize the value their distinctiveness provides to others.
Use and limitations
Instead of asking only for “more natural” prose, writers can specify vocabulary, sentence rhythm, regional expression, and clichés to avoid, then compare the revision with the source. Organizations should not confuse a consistent brand with identical voices.
This is a mathematical model with synthetic simulations, not a longitudinal measurement of real populations. It does not prove that a quantified monoculture has already occurred; validation across languages, cultures, and models is still required.