USC-led study in Nature Human Behaviour links LLM writing tools to declining linguistic diversity
A team led by the University of Southern California, with collaborators, has published a study in Nature Human Behaviour showing that the widespread use of large language models (LLMs) as writing assistants is associated with a decline in linguistic diversity.
Analyzing more than 880,000 texts—including news articles, academic papers, argumentative essays, social media posts, and speeches—the researchers found that LLM-assisted writing is linked to homogenized writing styles and a reduction in textual cues related to personal identity, personality, and mental health.
The study comprised three analyses. First, examining longitudinal datasets from arXiv, Patch News, and Reddit between 2018 and 2024, the team observed significant drops in writing complexity variance after ChatGPT’s release, with AI usage predicting subsequent variance changes. Controlled experiments also showed that LLM rewriting compressed stylistic variance. Second, using paired psychological measurement data, classifiers predicted personal traits from original versus LLM-rewritten texts. Prediction accuracy fell by about 6% on average, with the F1 score for age dropping from 0.351 to 0.260, and rewrites systematically skewed toward certain profiles. Third, several classic word–person associations disappeared after rewriting, such as extraversion and pronoun use, loyalty and friend-related words, and age and future-oriented vocabulary, while others persisted—indicating selective erosion.
The authors note that reduced linguistic diversity could complicate mental health screening, personalized marketing, fair hiring, and cultural preservation. They also highlight a potential feedback loop, as LLM-generated text entering online corpora may further reduce diversity in future models.
