LLM-Edited Text Forms a Cluster Separate from Human Writing

Researchers from UC Berkeley, UCSD, University of Washington, and Google DeepMind analyzed a 100-user experiment, 86 human essays from 2021, and 18,000 ICLR 2026 peer reviews.

The key finding is that essays edited by LLMs clustered in a separate region of embedding space that human writing does not occupy. Even when the instruction was "just fix the grammar," the direction of the shift was nearly identical, and each writer's distinctive vocabulary traces were overwritten with LLM-preferred expressions.

• LLM-assisted writing was more neutral and took fewer definitive stances.
• Use of nouns and adjectives increased while pronouns decreased, shifting the style to be more formal and statistical.
• Arguments based on personal experience shifted toward ones centered on statistics and expert citations.

In the ICLR 2026 peer review analysis, 21% were estimated to be AI-generated, and AI reviews gave papers scores that were 10% higher. AI reviews were also 136% more likely to emphasize reproducibility and 84% more likely to emphasize scalability. Human reviews, on the other hand, mentioned clarity more often in both strengths and weaknesses.

The study shows that while LLM assistance increases efficiency, it can simultaneously alter both the individuality of writing and the criteria used for evaluation.