How LLMs Distort Our Written Language

LLMs increase writing efficiency, but when they revise human writing, they show a tendency to also change the conclusion, stance, and even type of argument.

The core findings fall into three strands.

• In a user study, people who used LLMs heavily reported being satisfied with the results, yet also said their voice and creativity had diminished — showing satisfaction and loss of expression occurring at the same time.
• When 86 human-written essays from ArgRewrite-v2 were edited by gpt-5-mini, gemini-2.5-flash, and claude-haiku, even instructions limited to minimal edits or grammar fixes produced significant changes in vocabulary and meaning.
• An analysis of 18,000 peer review comments from ICLR 2026 found that 21% were judged to be AI-generated, and these reviews gave scores 10% higher than human reviews, while placing greater weight on reproducibility and scalability.

Stylistic changes were also pronounced. Writing composed or edited by LLMs was more formal and impersonal, with an increase in nouns and adjectives and a decrease in pronouns. Arguments based on personal experience weakened, while emotional, logical, and statistical language and citations of expert opinion increased.

Semantic space analysis also showed that LLM-edited versions consistently shifted toward directions that human writing did not predominantly occupy. Even given the same prompt, they converged in different directions than human edits, confirming an effect that goes beyond simple correction to rearranging the meaning of the writing itself.

The study suggests that AI could subtly alter politics, culture, science, and even everyday messaging. Users are likely to keep choosing AI writing without fully recognizing these changes, and even in professional contexts, there remains an incentive to submit AI-generated text as one's own writing.