LLM Writing Assistants Flatten How We Write — and Blur Who Wrote It

A study by Sourati and colleagues in Nature Human Behaviour tracks what happens to prose once large language models become the default writing aid. Across three studies spanning seven datasets and more than 880,000 texts, the authors find that heavy LLM use correlates with measurable declines in linguistic diversity. Machine-polished writing converges toward a common style: variance in writing complexity drops by a statistically significant 21–50% depending on the dataset and model. Because LLMs are optimized to produce the statistically likeliest continuation of a text, they amplify dominant language patterns and suppress less common ones, nudging everyone toward the same middle-of-the-road register.

The more striking finding is what gets stripped out. When an LLM rewrites or ‘improves’ a passage, it tends to erase the subtle linguistic cues that signal an author’s gender, age, ideology, and moral values. The individual fingerprint in someone’s writing — the markers that encode identity and social context — flattens into a generic voice. Homogenization is not just an aesthetic concern; it removes information that downstream systems and readers actually rely on.

The authors argue the consequences reach well beyond style. Losing these identity signals could compromise processes that depend on them, from clinical or diagnostic assessment to personalization, and could deepen existing inequities — for example, in hiring and personnel selection, where machine-normalized text may disadvantage those whose natural voice diverges from the dominant pattern. They also warn of a slower erosion of cultural and linguistic heritage as global writing collapses toward a single statistical norm. The work reframes AI writing tools not as neutral polishers but as a homogenizing force with equity and cultural-preservation stakes.