The words people choose, the rhythm of their sentences and the complexity of their writing can reveal far more than what they intend to communicate. Language carries traces of identity, emotion, education, social position and psychological state, making it a powerful source of information for researchers, clinicians, employers and businesses. But a new study suggests that this information may be quietly eroding as large language models become routine writing assistants. When people ask an artificial intelligence system to polish, rewrite or improve their text, the result may be clearer and more fluent, yet also less distinctive. Across three studies involving seven datasets and more than 880,000 texts, researchers report that LLM-assisted writing is associated with a measurable narrowing of linguistic diversity.
The study, published in Nature Human Behaviour, examined how language changes when written material is revised by different large language models in a range of domains. The researchers focused not only on whether the central message survived the rewriting process, but also on what happened to the style surrounding that message. Their findings indicate that LLMs generally preserve the core content of a text while making the language more similar to other AI-assisted writing. This process, described as linguistic homogenization, reduces the variation that normally exists between people’s writing styles. Across datasets and models, the researchers observed statistically significant reductions of between 21 and 50 percent in the variance of writing complexity, with reported probability values of P ≤ 0.05.
Writing complexity is not a single feature. It can include sentence length, syntactic structure, vocabulary, lexical diversity and the way ideas are organized across a passage. Human writers naturally differ on all of these dimensions. One person may use short, direct sentences, while another may favor layered clauses, unusual metaphors or highly specialized vocabulary. These differences are not merely decorative. They can provide clues about the writer’s background, cognitive approach, emotional condition and relationship to the audience. When an LLM rewrites a text according to generalized expectations of clarity and professionalism, some of those individual signals may be softened or removed, even when the argument itself remains intact.
The researchers also found evidence that LLMs can amplify linguistic patterns associated with dominant characteristics while suppressing less common ones. In practical terms, an AI system may make writing conform more closely to the style it has learned to recognize as standard, acceptable or effective. That tendency can be useful when a writer needs a formal report, a concise application or a grammatically polished message. However, it may also cause distinctive voices to move toward the center of a statistical distribution. Expressions linked to minority perspectives, unconventional communication styles or culturally specific ways of speaking could become less visible when repeatedly filtered through systems optimized to generate broadly acceptable text.
This effect matters because language is increasingly used as an indirect measurement tool. Psychologists analyze word choice and sentence patterns to study personality, mental health and social behavior. Healthcare researchers investigate written language for possible signs of cognitive decline, depression or other conditions. Marketing teams use language to infer preferences and tailor messages, while employers may evaluate written responses during recruitment. If AI-assisted writing systematically changes the linguistic features on which these judgments depend, the apparent signal may no longer come entirely from the person being assessed. It may instead reflect the conventions of the model that edited the text.
The consequences could be especially significant in systems designed to personalize decisions. An algorithm attempting to infer an individual’s needs from a message might interpret standardized AI language as evidence of a shared psychological or social profile. A clinician could receive a polished account that obscures the patient’s natural way of expressing distress. A hiring manager might compare applicants whose writing has been normalized by similar tools, mistaking model-generated fluency for equivalent communication ability. In each case, the technology could improve surface readability while reducing the variation needed to distinguish one person, group or cultural context from another.
The findings also raise a broader question about feedback loops in the language ecosystem. Large language models learn statistical patterns from vast collections of human-produced text, and users increasingly return AI-generated or AI-edited material to the public information environment. As more writing is produced through similar systems, future models may encounter an even more uniform linguistic landscape. This could reinforce the same stylistic preferences over time, creating a cycle in which language becomes increasingly optimized for machine-recognized clarity and increasingly distant from the full range of human expression. The study does not establish that such a cycle will occur everywhere, but its results suggest that the conditions for it are already visible.
Importantly, the researchers report that the pattern remained across different models, prompts and contexts, indicating that it was not limited to one particular chatbot or instruction. That consistency strengthens the case that the phenomenon reflects a broader property of AI-assisted rewriting rather than an isolated software defect. At the same time, the findings do not mean that every use of an LLM destroys individuality, nor that human writing is always more informative or authentic. Instead, they point to a trade-off: automated editing can deliver speed, grammatical accuracy and conventional readability, but those benefits may come with a reduction in stylistic variation. Recognizing that trade-off will be essential as AI writing tools move from optional conveniences into everyday infrastructure.
The study’s central warning is therefore not that people should stop using language models, but that society should be more careful about what is lost when communication is standardized. Researchers may need to distinguish original writing from AI-assisted writing when interpreting linguistic data. Healthcare and hiring systems may require safeguards against treating model-shaped language as a direct reflection of a person’s traits. Developers could explore tools that preserve unusual phrasing, culturally specific expression and deliberate stylistic variation instead of automatically pushing every text toward the same ideal. As LLMs become invisible collaborators in emails, applications, medical histories and public discourse, the diversity of human language may depend on whether convenience is allowed to become conformity.
Subject of Research: The impact of large language model writing assistance on linguistic diversity, writing complexity and the preservation of individual and socially informative language patterns.
Article Title: The shrinking landscape of linguistic diversity in the age of large language models
Article References: Sourati, Z., Karimi-Malekabadi, F., Ozcan, M. et al. “The shrinking landscape of linguistic diversity in the age of large language models.” Nature Human Behaviour (2026). https://doi.org/10.1038/s41562-026-02550-0
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41562-026-02550-0
Keywords: large language models, LLMs, linguistic diversity, writing assistants, language homogenization, writing complexity, artificial intelligence, human expression, psychological assessment, cultural preservation
