Large language models (LLMs) are used by over a billion people globally, most often to assist with writing. In this work, we demonstrate that LLMs not only alter the voice and tone of human writing but also consistently alter the intended meaning. First, we conduct a human user study to understand how people actually interact with LLMs when using them for writing. Our findings reveal that extensive LLM use led to a nearly 70% increase in essays that remained neutral in answering the topic question. Significantly more heavy LLM users reported that the writing was less creative and not in their voice. Next, using a dataset of human-written essays that was collected in 2021 before the widespread release of LLMs, we study how asking an LLM to revise the essay based on the human-written feedback in the dataset induces large changes in the resulting content and meaning. We find that even when LLMs are prompted with expert feedback and asked to only make grammar edits, they still change the text in a way that significantly alters its semantic meaning. We then examine LLM-generated text in the wild, specifically focusing on the 21% of AI-generated scientific peer reviews at a recent top AI conference. We find that LLM-generated reviews place significantly less weight on clarity and significance of the research, and assign scores that, on average, are a full point higher. These findings highlight a misalignment between the perceived benefit of AI use and an implicit, consistent effect on the semantics of human writing, motivating future work on how widespread AI writing will affect our cultural and scientific institutions.
Figures & tables
Figure 1: LLM-generated revisions display a larger and more consistent semantic shift than human-written revisions of the same essays. Each point pair represents an essay from the ArgRewrite-v2 dataset before (D1) and after (D2) revision, embedded using MiniLM-L6-v2 ( Reimers and Gurevych, 2019b ) sentence embeddings and projected into two dimensions via PCA, a common approach for analyzing semantic differences ( Dhillon et al., 2015 ) . The top-left panel (grey) shows human revisions, while the remaining panels show revisions produced by prompting the LLM with human-written expert feedback and different edit instructions (see titles). Even the instruction to make minimal edits shows large shifts (top-right). Arrows indicate direction and magnitude of semantic change. Human revisions exhibit smaller and more diverse shifts, whereas LLM revisions produce large semantic shifts strongly aligned in a common direction, and to a region of space not previously occupied by any human-written essay.
Figure 2: Example of substantial argumentative rewriting by an LLM on essays from ArgRewrite-v2. Red text highlights segments removed or substantially reframed by the model, while green text highlights segments added to the human draft. The figure illustrates how LLM edits frequently alter the person’s intended conclusions, removing content that makes a particular claim, and editing the essay to be more neutral or positive about the technology of self-driving cars. Further, LLM edits will often remove human colloquialisms, anecdotes, or examples, leading to repetitive writing that loses the person’s voice.
Figure 3: Unique words in human-edited texts (left) versus AI-edited texts produced using gpt-5-mini (right) for the ArgRewrite-v2 analysis. Word size reflects relative frequency, highlighting stylistic and thematic differences between how humans and AI write about self-driving cars.
Figure 4: In a randomized controlled trial, users which engaged in heavy LLM-use report essays are less creative and not in their voice, and exhibit large, homogenizing semantic shifts in their writing.
Figure 5: Semantic shifts induced by human and LLM revisions for the ArgRewrite-v2 dataset. Each point pair represents an essay before (D1) and after (D2) revision, embedded using gemini-004 sentence embeddings and projected into two dimensions via PCA, a common approach for analyzing semantic differences ( Dhillon et al., 2015 ) . The left panel shows human revisions, while the remaining panels show revisions produced by different LLMs without access to expert feedback. Arrows indicate the direction and magnitude of semantic change. Human revisions exhibit smaller, more varied semantic shifts, whereas LLM revisions produce larger shifts that are strongly aligned in a common direction, indicating a homogenization effect in semantic space.
Figure 6: The results of using LLM-as-a-Judge to categorize the argumentative stance of essays written during the human user study. Relative to the control group, our results show that extensive AI use results in a 68.9% increase in the proportion of essays that remain neutral when answering the topic question, rather than actually expressing a for/against opinion ( t(69)=−2.439,p=0.017 ) comparing human to LLM responses.
Figure 7: Lexical divergence distribution for general revision of claude-haiku with expert feedback for ArgRewrite-v2. Gray bars show the human baseline, while colored distributions show LLM revisions systematically shift rightward with a mean distance of 0.476. Higher JSD indicates more extensive vocabulary replacement, with LLMs substituting more words than humans. Note that claude-haiku resulted in the least semantic distortion.
Figure 8: Parts of Speech Distribution. Left: Essays written during the user study, either by people in the control group, or with assistance from LLMs. Right: Essays from the ArgRewrite-v2 dataset edited by humans versus edited by 3 different LLMs prompted to make general edits. Across both studies, LLMs use more nouns and adjectives when writing/editing, and reduce the use of pronouns.
Figure 9: Emotional shifts for general revisions with and without expert feedback. Human baseline (gray) shows minimal affective changes. LLMs increase positive sentiment (37-54%) and trust-related language (17-53%), while also simultaneously increasing negative sentiment (24-38%). This indicates increases in the use of emotional language, regardless of which LLM model is used.
Figure 10: LLM assistance shifts writing to focus on analytical and statistical arguments (a) LIWC analytic scores are elevated in LLM-edited text. (b) Human study participants with LLM access rely more heavily on expert opinion ( t(69)=−2.78,p<0.01) and statistical arguments ( t(69)=4.82,p<0.0001 ).
Figure 11: We use LLM-as-a-Judge on 18k reviews from ICLR 2026, selected because they were all written on the subject of LLMs, where 9k reviews are entirely written by humans, and 9k are entirely written by LLMs. We plot the relative frequency of certain strengths and weaknesses occurring in LLM reviews instead of human reviews. For the Clarity strength category, LLMs were 32% less likely to select it compared to a baseline of 17% in human reviews. A two-proportion z-test confirmed this difference is highly significant ( z=11.00,p<0.001 ), indicating a systematic divergence in how AI models prioritize presentation-based critiques compared to human peers.
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12: We report the demographic statistics for ethnicity, sex, and student status. Some participants do not report their demographics in certain categories, so it does not all add up to 100 users.
Figure 13: Semantic shifts from D1 to D2 for General revisions with Gemini-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 14: Semantic shifts from D1 to D2 for Grammar edits with with Gemini-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 15: Semantic shifts from D1 to D2 for Miminal revisions with Gemini-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 16: Semantic shifts from D1 to D2 for Completion revisions with Gemini-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 17: Semantic shifts from D1 to D2 for Expansion revisions with Gemini-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 18: Semantic shifts from D1 to D2 for General revisions with MiniLM-L6-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 19: Semantic shifts from D1 to D2 for Grammar edits with MiniLM-L6-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 20: Semantic shifts from D1 to D2 for Miminal revisions with MiniLM-L6-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 21: Semantic shifts from D1 to D2 for Completion revisions with MiniLM-L6-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 22: Semantic shifts from D1 to D2 for Expansion revisions with MiniLM-L6-004 Embedding . Top: with expert feedback. Bottom: without expert feedback.
Figure 23: Semantic shifts from D1 to D2 produced by GPT-5-mini . Top: revisions with expert feedback. Bottom: revisions without feedback.
Figure 24: Semantic shifts from D1 to D2 produced by Gemini-2.5-Flash . Top: revisions with expert feedback. Bottom: revisions without feedback.
Figure 25: Semantic shifts from D1 to D2 produced by Claude-3.5-Haiku . Top: revisions with expert feedback. Bottom: revisions without feedback.
Figure 26: Lexical divergence (JSD) distributions for General revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 27: Lexical divergence (JSD) distributions for Grammar revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 28: Lexical divergence (JSD) distributions for Minimal revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 29: Lexical divergence (JSD) distributions for Completion revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 30: Lexical divergence (JSD) distributions for Expansion revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 31: Grammatical shifts (POS) for General revisions . Top: with expert feedback. Bottom: without feedback.
Figure 32: Grammatical shifts (POS) for Grammar revisions . Top: with expert feedback. Bottom: without feedback.
Figure 33: Grammatical shifts (POS) for Minimal revisions . Top: with expert feedback. Bottom: without feedback.
Figure 34: Grammatical shifts (POS) for Completion revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 35: Grammatical shifts (POS) for Expansion revisions . Top: with expert feedback. Bottom: without expert feedback.
Figure 36: Emotional shifts from D1 to D2 for General revisions . Left: with expert feedback. Right: without expert feedback.
Figure 37: Emotional shifts from D1 to D2 for Grammar revisions . Left: with expert feedback. Right: without expert feedback.
Figure 38: Emotional shifts from D1 to D2 for Minimal revisions . Left: with expert feedback. Right: without expert feedback.
Figure 39: Emotional shifts from D1 to D2 for Completion revisions . Left: with expert feedback. Right: without expert feedback.
Figure 40: Emotional shifts from D1 to D2 for Expansion revisions . Left: with expert feedback. Right: without expert feedback.
Strength Dataset
Category
G1 %
G2 %
% Change
Z-Score
P-Value
Clarity
5.76
3.89
-1.87
14.3704
0.0000
Practical Application
2.02
3.42
1.40
-14.1956
0.0000
Reproducibility
0.62
1.46
0.84
-13.6601
0.0000
Relevance
3.45
2.35
-1.10
10.7794
0.0000
Theoretical Foundations
3.10
4.02
0.92
-8.1627
0.0000
Appendix
Table 1: G1 and G2 denote the proportional distribution of review categories. Because the methodology allows for a maximum of three strength and three weakness labels per entry, the denominator is normalized by tripling the total number of reviews to reflect the maximum potential labels.
The number of paper submissions to major AI conferences has increased dramatically in recent years, substantially increasing the peer-review workload of researchers. At the same time, large language models (LLMs) are increasingly used to support peer review, both by individual reviewers and by conference organizations. While LLM capabilities have improved considerably, their evaluations may be influenced by linguistic characteristics that are unrelated to scientific content. In this work, we introduce a controlled experimental framework for studying such effects by systematically perturbing the linguistic style of research papers. Specifically, we rewrite papers to exhibit (i) linguistic features associated with native or non-native English writing, and (ii) high- or low-confidence linguistic features. To construct these interventions, we derive fine-grained rewriting instructions grounded in the linguistic literature and apply them using LLMs. Crucially, our interventions are designed to preserve the underlying scientific content. We then evaluate the resulting papers using different LLM reviewers. Our experiments expose a critical vulnerability in LLM-based peer review: reviewers systematically reward confident rhetoric independent of scientific content, raising serious concerns about the robustness of their assessments.
Sai Suresh Macharla Vasu, Sumedh Santosh Joshi, Philipp Christmann +6
CISPA Helmholtz Center for Information Security · German Research Center for Artificial Intelligence (DFKI) · University of Technology Nuremberg +1
Has the style of scientific communication changed due to the growing use of large language models in the writing process? We address this question in the domain of Natural Language Processing by leveraging two data resources we create: a naturalistic corpus of over 37,000 papers from the ACL Anthology (2020-2024); and a synthetic dataset of 3,000 human-written passages and their LLM-generated improvements. We first implement a series of diachronic lexical analyses, showing that both word frequency and usage contexts have changed significantly over time, indicating semantic specialization in some cases and generalization in others. Broadening our perspective, we then model a range of more complex stylistic features and find that LLM-modified texts more frequently contain certain syntactic constructions, more complex and longer words and a lower lexical diversity. Finally, we connect these changes in writing practices to subjective reading experience through a pilot annotation study with 20 domain experts. They overall rate LLM-improved texts as more understandable and exciting, but also express negative qualitative attitudes towards LLMs, highlighting the strongly subjective effect of AI-assisted writing on reading experience.
Filip Miletić, Neele Falk
Institute for Natural Language Processing, University of Stuttgart, Germany
Over the past several years, LLM-powered chatbots and agents have become widely used as a tool for academic writing. LLM-assisted writing can be valuable by removing language barriers but at the same time causes concerns about misconduct and fraud. To inform policy decisions, it is necessary to monitor the prevalence of LLM-altered texts in scholarly publications. Despite some recent progress in this direction, no existing method can produce reliable estimates. Here we suggest and validate a new unbiased approach to estimate LLM usage in a corpus of texts based on changing word frequencies. We apply our method to the full texts of open-access biomedical papers from Pubmed Central, and show that by the end of 2025, 89% of papers show excess of LLM-associated vocabulary. We also find that LLMs are twice as likely to be used when writing a paragraph in the Discussion section (68%) compared to a paragraph in the Methods section (32%), but even inside the Methods section, the overall prevalence of LLM usage is over 50%. We believe that our estimates are crucial to shape future guidelines and policies.
Lena Holzwarth, Rita González-Márquez, Dmitry Kobak
1Hertie Institute for AI in Brain Health, University of Tübingen, Germany · Department of Mathematics, Computer Science, and Statistics, Ghent University, Belgium · 3VIB Center for AI and Computational Biology, VIB, Ghent, Belgium