Contribution
The paper puts numbers on a common impression: LLM help with writing changes what the writing says, not only how it sounds. It does so in three settings. There is a randomized user study of a short argumentative essay, and a counterfactual re-editing of a 2021 revision corpus whose human revisions predate ChatGPT. There is also a comparison of human and LLM-generated ICLR 2026 reviews. Its new element is the counterfactual: the same draft and the same expert feedback, revised once by its author and once by an LLM.
Key insight
When a person revises their own draft, the change in embedding space is small and points in a different direction for each essay. When an LLM revises the same drafts, the change is large and points the same way for every essay and every model, into a region no human essay occupies (Figures 1, 5). The distortion is a drift toward one shared place, not noise.
Assumptions
- Meaning is position in a sentence-embedding space. “Distances in our PCA plot reflect similarity in meaning or semantic content between essays” (§3.2). MiniLM-L6-v2 and gemini-004 embeddings are used, projected by PCA or t-SNE to two dimensions.
- Stance is what gpt-4o says it is. An LLM judge labels each user-study essay “for”, “against” or “neutral” on “Does money lead to happiness?” (Appendix D.2). No writer was asked their own view, and no human validated the labels.
- The human revision is the target. “[T]he true objective should be to produce text similar to what the human would have if they had invested the time and effort to do it themselves” (§5).
- Pangram’s labels are taken as ground truth for which ICLR reviews are LLM-generated. The authors note it “is possible for this tool to have errors and to introduce correlations” (§3.1, fn. 1).
- One topic per study. The user study has one essay prompt. ArgRewrite-v2 has one topic, self-driving cars, and 86 essays. The participants are US native English speakers on Prolific.
Key results
- User study, voice (§4.1, Fig. 4a). N = 100: 45 control, 55 with an embedded gpt-4o-mini. Of the 55, 28 were classed “LLM-Influenced” (less than 40% of the text LLM-generated, checked against transcripts) and 27 “LLM”. The heavy users rated their essays less creative (t(69) = 2.110, p = 0.0385) and less “in my voice” (t(69) = 5.799, p < 0.001). Satisfaction, difficulty organizing and struggle did not differ.
- User study, stance (§4.3, Fig. 6). The judge labelled 39.5% of control essays neutral and 66.7% of heavy-LLM essays, “a 68.9% increase in the proportion of essays that remain neutral”. The caption gives t(69) = −2.439, p = 0.017, and the text “p < 0.036”. The LLM-Influenced group stayed close to the control in stance, in embedding space and in self-report.
- ArgRewrite-v2, semantic shift (§4.2, Figs. 1, 5, 13–25). The LLMs were gpt-5-mini, gemini-2.5-flash and claude-haiku. Human D1→D2 shifts are “smaller and more diverse”. LLM shifts are “large semantic shifts strongly aligned in a common direction” in every edit mode, with and without the expert feedback, and under both embeddings. gpt-5-mini moves furthest and claude-haiku least. Completion and expansion move more than the lighter modes. The evidence is visual: no distance, angle or test statistic is reported for this result.
- ArgRewrite-v2, stance (§4.3, Fig. 2). This result is qualitative. In the paper’s grammar-edit example (claude-haiku), “since the perks do not outweigh the persistent cons, perhaps America is not ready” becomes “since the drawbacks do not outweigh the benefits, America is not ready”. The shared mode is described as “essays that are in support of self-driving cars”. The stance of the LLM-revised essays was not counted.
- Lexical (§4.4, Fig. 7). Human revisions keep the Jensen-Shannon divergence of their unigram distributions at about 0.2–0.3 from D1. claude-haiku general revisions average 0.476. gpt-5-mini reaches “nearly triple the human baseline”.
- Grammar and affect (§§4.5–4.7). Pronouns fall 40–60% and adjectives rise 57–90% across models. Positive sentiment rises 37–54%, negative 24–38% and trust 17–53% (NRC lexicon). Human edits change each category by under 5%. LIWC “analytic” scores rise. Heavy users’ essays cite expert opinion and statistics more (t(69) = 4.82, p < 0.0001).
- ICLR 2026 reviews (§4.8, Fig. 11, Table 1). About 9k fully human and 9k fully LLM reviews, labelled by a gpt-4o pipeline with up to three strengths and three weaknesses each. Mean score: 4.43 for LLM reviews against 4.13 for human reviews. Clarity as a strength: 5.76% to 3.89% of label slots (−32%, z = 14.37). Clarity as a weakness: 4.62% to 1.93%. Relevance as a strength: 3.45% to 2.35%. Reproducibility as a strength rose 0.62% to 1.46% (+136%). Impact as a strength fell 1.05% to 0.82%.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Heavy LLM users judge their essays less in their own voice and less creative, but are as satisfied | moderate | t-tests on Likert items, n = 27 against 45, §4.1 |
| C2 | Heavy LLM use makes an essay more likely to take no side | weak | one prompt, an unvalidated LLM judge, n = 27 against 43, two p-values that disagree, §4.3 |
| C3 | LLM revisions move essays further, and in a common direction, in embedding space than authors’ own revisions | moderate | 86 essays, three models, five prompts, two embeddings, all shown as plots without statistics, §4.2 and Appendix B |
| C4 | LLM revisions change an essay’s conclusion even under a grammar instruction | weak | worked examples (Fig. 2, Appendix I), no count |
| C5 | LLM-generated reviews apply different criteria from human ones | moderate | 18k reviews, z-tests on judge labels, Table 1; rests on Pangram’s labels |
| C6 | LLM revisions shift the author’s intended meaning | weak | inferred from C2–C4; intended meaning is never measured |
Method
- User study. Prolific participants wrote on “Does money lead to happiness?” for $8 in up to one hour, with pre- and post-questionnaires (Appendix A). Group membership among LLM users was fixed “a priori” from self-report checked against transcripts.
- Counterfactual editing. Each ArgRewrite-v2 D1 draft was given to
three LLMs with the expert’s feedback, which covers claims, evidence and
reasoning as well as surface errors. The five prompts are in Appendix B.1:
- Grammar: “Revise this essay for grammar.”
- Minimal: “Rewrite this essay and keep to a similar word count.”
- Expert (general): “Revise the essay.”
- Completion: “The text is only the first paragraph. Finish the rest of the essay.”
- Expansion: “Expand on the ideas in the following draft with more detail and depth.” The outputs are compared with the human D2 on embeddings, JSD, POS, NRC and LIWC. A second run omits the feedback.
- Reviews. A three-stage gpt-4o pipeline (Appendix D.1) builds at most 15 categories, merges them, and then assigns at most three strengths and three weaknesses to each review. Proportions are compared by two-proportion z-tests.
Concepts
- Distortion. A change in meaning, stance, voice or evaluative criteria that the writer did not ask for. In practice it is measured as distance and direction in embedding space, lexical divergence, and shifts in judge labels.
- LLM-Influenced against LLM. Users who consulted the model, for under 40% of the text, against users who let it generate the text.
- Counterfactual edit. The revision an author would have made without an LLM, taken as the human D2 written in 2021.
Connections
The paper places itself in the homogenization literature (Jiang et al. 2025, the “artificial hivemind”; Doshi and Hauser 2024; Agarwal et al. 2025 on Westernizing suggestions). It also cites Qiu et al.’s lock-in hypothesis. The ICLR detection figures are Pangram’s (Emi 2025). It explains the effect, speculatively, by RLHF optimizing for “broadly more convincing” text, and compares that with engagement-optimizing recommenders.
Bearing on the record
- For QUESTION-002. It is a measured case of a rendering, an edit or a co-written text, that keeps the topic and the genre and shifts stance, voice and evaluative emphasis. It does not supply what QUESTION-002 asks for. Its measure of meaning is semantic similarity in a sentence-embedding space, the measure QUESTION-002 says an answer must beat. It collects no judgements of fidelity from the writers or from readers. So it shows that embedding drift and judge-labelled stance move together under LLM editing. It cannot show that either tracks what the writer meant.
- The LIT’s filing text overstates two results.
- “changed their meaning even when asked only for grammar edits”: the grammar prompt is “Revise this essay for grammar”, given in the main condition together with expert feedback on claims, evidence and reasoning. The “minimal” prompt asks for a rewrite at a similar word count, not for minimal edits. The grammar-mode shift is shown in plots and examples, not tested. The paper’s own abstract uses the phrase “asked to only make grammar edits”.
- “LLM-written peer reviews weighted clarity and significance less”: the body says clarity, relevance and impact. The abstract’s “scores that, on average, are a full point higher” is not the body’s 4.43 against 4.13, a difference of 0.3 on ICLR’s 0–10 scale, which the body calls “10% higher”.
- Internal inconsistencies to know before citing.
- Two p-values for the stance result (0.017 and ”< 0.036”).
- The claude-haiku model is named as “Claude-3.5-Haiku” (Fig. 25) and as “Claude Haiku 4.5” (Appendix I).
- The ICLR sample is described two ways: as papers with one wholly human and one wholly LLM review (§3.1), and as reviews “selected because they were all written on the subject of LLMs” (Fig. 11).
- For CLAIM-115. The heavy-user condition gives a rough test: the receiving system (gpt-4o as reader) no longer recovers a side on the question asked. That fits a relational account of what is lost better than a lexical one. But the measure is a single coarse label.
- For CLAIM-147’s Study B. ArgRewrite-v2 with this paper’s five prompts is a ready corpus of graded transformations of the same source. Each comes with a human counterfactual, so whether relational measures beat surface similarity could be tested on it.
- No instruction for machine-learning practice is carried as a finding. The authors’ call for models that infer intended meaning is a research direction. It is close enough to ML practice that the anthology may want the paper for its own question (ADR-013); it is not held there now.
Limitations
- Intended meaning is never measured. “These results imply that the LLM-generated essays fail to convey users’ actual opinions” (§4.3) infers it from a group difference in judge labels.
- One essay prompt in the user study and one topic in the corpus. The authors say the results “may present themselves differently in different languages or for people residing in different countries”.
- The heavy-user group is self-selected within the randomized arm. The stance contrast compares it with the control arm, so it is not a randomized comparison.
- The main semantic result has no effect size or test. The paper does not say which essays the PCA was fitted on. If it was fitted on human and LLM essays together, the first component will tend to separate them, and a common LLM direction is partly built into the projection.
- The ICLR comparison depends on Pangram’s classifier and on gpt-4o’s categories. Many of the percentage changes are on base rates of 1–2% of label slots.
Open questions
- Do the writers themselves judge the LLM revision as saying something different? An author-rated fidelity study on ArgRewrite-v2 would close this, and would also test embedding distance as a proxy.
- How much of the grammar-mode shift comes from the expert feedback’s content instructions? A grammar-only prompt with no feedback, scored for stance, would separate them. The appendix plots exist, but no numbers.
- Is the common direction a property of the topic (pro-technology on self-driving cars) or of the models? It would need a corpus with several topics.