How LLMs Distort Our Written Language

Search this site

Embedded Files

Skip to main content

Skip to navigation

How LLMs Distort Our Written Language

Marwa Abdulhai,  Isadora White, Yanming Wan, Ibrahim Qureshi,<br>Joel Z. Leibo, Max Kleiman-Weiner, Natasha Jaques<br>UC Berkeley,  UC San Diego, University of Washington, Zaytuna College, Google DeepMind<br>Paper | Code

Executive Summary<br>LLMs are used by over a billion people globally, and the most frequent use case is to assist with writing. LLMs can provide a huge efficiency boost, but are they actually writing what we want?<br>Many users recognize the "feel" of LLM prose, but few people realize the extent to which LLMs distort the meaning of writing. We find this across three datasets: a human user study, a dataset of human argumentative essays, and reviews from a top machine learning conference.

Main Findings<br>LLMs change the conclusions of writing, changing the stance as well as the argument type

Human users report a paradox of preferences, being satisfied while reporting a statistically significant loss of voice and creativity

LLMs introduce larger semantic shifts than human edits do, even when prompted only to introduce grammar edits

These shifts apply even to our institutions: LLM reviews gave significantly different reasons for acceptance and rejection at the International Conference of Learning Representations (2026), a top AI conference, the 21% of peer reviews that were found to be AI-generated focused on significantly different scientific criteria.

Why should we care?<br>As LLMs are integrated into society, these subtle changes in meaning could fundamentally alter politics, culture, science, and even the way we communicate with our friends and family. Our study focuses on argumentative writing, but our findings may generalize to many other forms of writing and communication as well.

When LLMs revise human writing, they induce large homogenizing changes very unlike how people would have edited the same essay.<br>This figure shows a counterfactual analysis, comparing how LLMs edit a piece of text to how a person would. The upper left grey figure shows how people make edits to a first draft essay when they receive expert feedback. The first draft is shown as a light grey dot, with an arrow to the dark grey second draft, visualized using the MiniLM-L6 semantic embedding space and projected with PCA. The remaining panels show how an LLM edits the original human-written essay, when it is prompted with the expert feedback and a series of prompts. Even when the LLM is instructed to make minimal edits, we see that it induces large changes to all the essays, moving them in a consistent direction away from how humans write.

Above is an intuitive an example of how writing with an LLM alters the conclusions and removes the humans voice in essays from ArgRewrite-v2 dataset.<br>We illustrate this more extensively in our results below.

Methodology & Datasets

We study how LLMs distort meaning in our written language in three datasets.

Human User Study: To understand how humans use LLMs while writing, we conduct a human user study, with 55 users enabled to use the LLM and 45 without access to the LLM. Since many human users chose to abstain from LLM use during their session, we condition our results on this choice and split into two groups: LLM-Influenced, for those who chose not to use or use only for information seeking, and LLM, the group of extensive users. We split them into these groups a priori, by observing their transcripts, final essays, and their self-reported usage score, before evaluating and running our analyses.

ArgRewrite-v2: Using a dataset of 86 human-written essays collected in 2021 — before the widespread release of LLMs — we prompted three production LLMs (gpt-5-mini, gemini-2.5-flash, claude-haiku) to edit essays across five revision types: general revision, minimal edits, grammar edits, completion, and expansion. We compare LLM-generated drafts to human-written revisions along dimensions of semantics, lexical usage, part-of-speech distributions, emotional tone, and stylistic features.<br>ICLR 2026 Review Analysis: We analyze 18k peer reviews from ICLR 2026, selecting papers with one entirely human-written review and one entirely LLM-generated review. We use an LLM-as-a-Judge classifier to identify the strengths and weaknesses cited in each review and compare scores assigned by humans vs. LLMs.

Heavy LLM users report that their essays do not reflect their own voice.<br>This presents a paradox of preferences where the user reports satisfaction, but report a significant loss in creativity and voice.<br>RLHF optimizes for preferences, but this is not sufficient for maintaining creativity and semantics.

LLMs distort writing by shifting essays in a common semantic direction.<br>Essays written by humans in the control group are widely spread out throughout the embedding space, occupying a broad region that reflects the diversity of individual perspectives, writing...