Overview
- Researchers combined large observational datasets and controlled tests to measure the effect of LLMs on writing across academic papers, news articles, and social posts.
- Controlled rewrites by GPT‑3.5, Gemini, and Llama 3 cut variation in writing complexity by roughly 21% to 50% while keeping the original meaning.
- Statistical models lost about 6% average accuracy at predicting authors’ demographics and personality after texts were polished by LLMs, showing subtle identity cues were smoothed away.
- The scale of the change matters because a large share of new web content is produced with AI tools, which can multiply any trend toward uniform language styles.
- Authors and researchers are urging that model design prioritize stylistic diversity and that educators teach independent reasoning so students do not adopt homogenized writing as their default.