How LLMs Distort Our Writing

LLMs, used by more than 1 billion people worldwide, are most commonly used for writing assistance. But the results show that editing goes beyond grammar correction to changing the meaning, argument, and voice of the writing.

In the MiniLM-L6 embedding space, human-edited writing spread widely, while LLM-edited versions converged into a narrow region rarely occupied by human writing. Even when editing the same sentences, LLMs pushed the writing in a different direction than human edits, flattening out meaning.

• Human user study: 55 people were allowed to use LLMs while 45 people were not, and based on actual usage levels they were split into LLM-Influenced and LLM groups. Heavy users were satisfied with the results, but a preference paradox was confirmed—their own voice and creativity had diminished.
• ArgRewrite-v2: 86 human essays collected in 2021 were edited by gpt-5-mini, gemini-2.5-flash, and claude-haiku using general revision, minimal revision, grammar correction, completion, and expansion methods. Even when instructed to fix only grammar, a much larger semantic shift occurred compared to human editing, and the writing shifted toward being more neutral, formal, and analytical. Statistics, logic, and expert citations increased over personal experience, and both positive and negative emotional language increased as well.
• ICLR 2026: An analysis of 18,000 reviews found that about 21% were identified as AI-generated. LLM reviews scored about 10% higher than human reviews, and while humans more often mentioned clarity and research relevance, LLMs emphasized reproducibility and scalability more.

In the end, LLMs don't stop at polishing sentences—they can change conclusions and evaluation criteria as well. While convenience has increased, this serves as a warning that unique perspectives in politics, culture, science, and everyday conversation could be diluted.