Artificial intelligence writing assistants are making people's writing styles more similar to one another, according to a study published in the journal Nature Human Behaviour. Researchers found that while large language models preserve the core content of a text when they refine or rewrite it, they also flatten out the individual style of the writer.
The study was led by the University of Southern California and analysed more than 880,000 texts, including stories from Reddit, press articles, academic papers, essays, social media posts and political speeches.
Language carries a great deal of information about a person's identity, psychological state or social context, the researchers noted. But they warned that the growing use of AI tools such as writing assistants is pushing styles closer together and producing common linguistic patterns.
How the researchers tested the models
The team asked three large language models, GPT-3.5, Llama 3 70B and Gemini Pro, to rewrite thousands of texts originally written by humans. They then compared the style of the original texts with the style of the AI-rewritten versions.
The results showed that variability in writing complexity fell by a statistically significant margin of between 21% and 50% across all the datasets and all three models tested.
The models tended to amplify patterns associated with dominant writing characteristics while suppressing others, the study found, effectively prioritising conformity over individuality.
Meaning preserved, style homogenised
Despite the shift in style, the substance of what was written largely survived the rewriting process. In 87% of cases, the original and rewritten texts scored above 0.95 on a measure of semantic similarity, indicating that the models preserved the original meaning to a large extent.
The researchers said the same tendency toward homogenisation appeared in every scenario in which large language models were used to improve or rephrase text, regardless of the type of writing involved.
They said this pattern could have consequences for fields including diagnostic processes, personalisation efforts, hiring evaluations and cultural preservation, since each of these can depend on picking up on distinctive features of a person's language.
Why researchers say it matters
Large language models, or LLMs, are AI systems trained on vast amounts of text that can generate, summarise or rewrite writing on demand. Tools built on them, such as chatbots and writing assistants, are increasingly used by professionals and casual writers alike to polish emails, essays, articles and other documents.
The researchers wrote that as AI-mediated communication becomes more widespread, it is essential to ensure these tools enhance, rather than extinguish, what they described as the rich tapestry of human linguistic diversity.
Impact on identifying personal traits
The study also tested how well computational models could identify personal characteristics of authors based on their writing after it had passed through an AI rewrite. Accuracy dropped by an average of six percentage points compared with analysis of the original, human-written texts.
The researchers said this happened because the rewriting process dulled some of the linguistic patterns that are normally associated with those personal traits.
They concluded that writing assisted by large language models could reduce the reliability of language-based assessments in fields such as psychology, mental health support, personnel selection and personalised services, where subtle cues in a person's writing are often used to draw conclusions about them.
