LLM-Assisted Writing Risks Linguistic Monoculture, Study Warns

As large language models (LLMs) become ubiquitous in drafting and refining text, there's a growing concern that widespread adoption could lead to a "linguistic monoculture," where the diversity of language across a population diminishes. This research introduces a mathematical framework to model how authors and LLMs interact and coevolve, impacting linguistic variation. The study analyzes three interaction mechanisms: a shared, fixed LLM; a shared LLM that updates recursively from user outputs; and personalized LLMs. It finds that shared models can push authors towards a common linguistic norm, while recursive feedback merely shifts this norm without increasing diversity. Personalized models, however, show potential for preserving distinct author-model equilibria and maintaining linguistic diversity. The research also considers conformity as a strategic choice, where individuals might over-conform to gain benefits like clarity, but at the cost of distinctiveness. This creates a "price of monoculture," a negative externality where individual rational choices lead to a socially suboptimal reduction in linguistic diversity.