Overview
- A Nature Human Behavior study finds large language models push written text toward common stylistic norms, narrowing how people express ideas across academic papers, news and Reddit posts.
- In controlled tests, GPT-3.5, Gemini and Llama 3 rewrites cut variation in writing complexity by about 21% to 50% while largely preserving original meaning.
- Polishing texts with LLMs reduced the accuracy of statistical models that infer authors’ demographics and traits by roughly 6% on average, with effects varying by trait.
- Pew Research data cited by the reporting shows more than one third of web pages published since ChatGPT’s late-2022 launch were authored by AI, increasing the reach of these stylistic shifts.
- Lead researchers warn the changes could erode cognitive and expressive diversity—students are singled out as especially vulnerable—and they urge model and policy changes to protect individual voice.