Linguistic Monoculture Formalized: Shared LLMs Converge Authors to a Common Norm, Personalization Preserves Diversity

arXiv 2607.27134·low signalXBlueskyLinkedInCopy link

arXiv 2607.27134 (2026-07-29) builds a mathematical framework in which authors and LLMs are distributions over linguistic features that coevolve through repeated interaction, then analyzes three regimes: a shared model with a fixed distribution, a shared model recursively updated from author outputs, and personalized models updated via author-specific and population-level feedback. Characterizing equilibria and convergence rates, they show shared models drive authors toward a common norm; recursive feedback relocates that norm without changing pairwise spread under common conformity; and only personalization preserves a family of distinct equilibria with nonzero linguistic diversity. It is theory without empirical measurement, but it gives a precise mechanism for why widely-shared writing assistants flatten prose.