Large Language Models and Shifts in Scholarly Writing Style: A Cross-Journal Quantitative Analysis of Ophthalmology Research Articles

Large Language Models and Shifts in Scholarly Writing Style: A Cross-Journal Quantitative Analysis of Ophthalmology Research Articles | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Large Language Models and Shifts in Scholarly Writing Style: A Cross-Journal Quantitative Analysis of Ophthalmology Research Articles Tom Kornhauser, Tolossa Tufa Regassa, Morris E. Hartstein This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9082140/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Large language models (LLMs) are increasingly integrated into scientific writing workflows, raising questions about whether their widespread availability may influence the language of the published scientific record. We conducted a longitudinal text analysis to examine whether stylistic features of research articles changed following the introduction of widely accessible LLM tools. A corpus of 862 full-length original research articles was assembled from four general ophthalmology journals representing Clarivate Journal Citation Reports quartiles Q1–Q4. Articles were sampled systematically by journal-month from pre-LLM (January 2018–December 2020) and post-LLM (January 2023–July 2025) periods. Using an automated text-processing workflow, we quantified lexical discourse markers and punctuation features associated with editorial and connective phrasing patterns in scientific writing. Feature frequencies were normalized by article length, and a composite stylistic divergence index was constructed using standardized feature values within each quartile. Post-LLM articles showed measurable stylistic shifts, most pronounced in Q3 and Q4 journals. Several discourse and editorial markers increased in prevalence, punctuation patterns shifted, and the composite stylistic divergence index increased significantly in lower-quartile journals while remaining stable in Q1. Explicit disclosure of generative tool use was rare, occurring in fewer than 3% of post-LLM articles. These findings suggest that corpus-level stylistic patterns in scientific writing may be evolving in the post-LLM era and illustrate how quantitative analysis of linguistic features can help monitor technological influences on scholarly communication. Large language models scholarly communication scientometrics text mining scientific writing bibliometrics Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Large language models (LLMs), such as ChatGPT, Claude, Gemini, and Grok, are increasingly being used to assist with drafting, revising, and refining scientific manuscripts. While these tools may improve efficiency, they may also alter how research is phrased, structured, and rhetorically framed, even when their use is not explicitly disclosed.(Bao et al. 2025 ; Desaire et al. 2024 ; Nguyen et al. 2024 ) Because scholarly communication depends on shared conventions of tone, clarity, and presentation, the growing use of LLM-assisted writing may influence linguistic patterns within the published scientific record. Most discussion of LLM use in research has focused on ethics, transparency, disclosure, and authorship.(Misra and Chandwar 2023 ; Resnik and Hosseini 2025 ; Stokel-Walker 2023 ) However, fewer studies have examined whether the widespread availability of these tools is associated with measurable changes in the language of published articles themselves. Quantitative analysis of corpus-level linguistic features may provide an indirect way to study evolving writing practices, including potential shifts that occur without explicit disclosure. A field-bounded publication corpus provides a useful setting for examining such changes. Journals within a single biomedical specialty often share similar article structures, editorial conventions, and genre expectations, while still differing in impact level, authorship composition, and publication culture. Ophthalmology offers a suitable case study because it combines structured research reporting with internationally distributed authorship and clearly stratified journal quartiles.(Guerin et al. 2009 ; Li and Ge 2009 ) Building on earlier analyses of linguistic shifts associated with generative language models,(Amirjalili et al. 2024 ; Bao et al. 2025 ; Ji et al. 2023 ) we analyzed full-length original research articles from four ophthalmology journals spanning quartiles Q1 through Q4. Rather than attempting to identify individual LLM-generated manuscripts, we focused on corpus-level changes in lexical, discourse, and punctuation features that may reflect evolving patterns of scientific writing. Our goal was to quantify how scholarly writing style may be changing in the post-LLM period and to assess whether these changes vary across journal quartiles, disclosure behavior, and authorship-affiliation patterns. Data and Methods We assembled a longitudinal full-text publication corpus of original research articles from four general ophthalmology journals spanning journal quartiles Q1 through Q4. Quartiles were defined using Clarivate Journal Citation Reports (JCR) journal quartile assignments based on Journal Impact Factor category ranking (accessed March 5, 2026). The included journals were Ophthalmology (Q1), Graefe’s Archive for Clinical and Experimental Ophthalmology (Q2), BMC Ophthalmology (Q3), and the Turkish Journal of Ophthalmology (Q4). To reduce confounding by subspecialty scope and heterogeneous formatting, we restricted selection within each quartile to high-volume, broadly scoped general ophthalmology journals that publish primarily in English and had consistent monthly publication with stable PDF article structure across both study eras. Within each quartile, we selected a single representative journal meeting these criteria to enable systematic journal-month sampling and to prioritize internal comparability within a bounded disciplinary corpus over breadth. The pre-LLM period was defined as January 2018 through December 2020 and the post-LLM period as January 2023 through July 2025. This temporal split was designed to capture writing trends before and after the widespread availability of generative LLMs, while omitting the transitional period of 2021–2022, during which LLM exposure was growing but adoption remained inconsistent. For each journal and calendar month, two to four eligible original research articles were systematically sampled. Reviews, editorials, letters, case reports and case series, brief communications, comments or perspectives, protocols, conference abstracts, non-English texts, and supplements or special issues were excluded. We used a custom automated workflow written in Python to extract text from PDFs, remove headers, footers, and reference sections, and compute linguistic features. This pipeline extracted article body text, removed recurrent boilerplate, truncated text at mechanically identifiable reference sections when present, and tokenized the remaining body text so features could be counted consistently across documents. The extractor used open-source utilities (pdfminer.six, PyPDF2), and article-level word counts were recorded for normalization so that all linguistic features could be expressed relative to article length rather than as raw counts. The workflow performed text extraction and feature counting only. It did not involve training or fine-tuning any machine-learning model on publisher content, and it did not submit article text to external LLMs. Analyses were performed on locally stored copies of articles that were accessed through institutional subscription access or open-access availability, and results are reported only in aggregate feature counts and rates rather than redistributing full text. No full-text content is reproduced in this manuscript or shared as part of the study outputs; only derived, non-reconstructive feature measures and aggregate results are reported. To assess explicit disclosure of generative-tool use, we searched the full text of articles from the post-LLM period for a prespecified list of model and tool terms and common variants. Search terms included ChatGPT, GPT, GPT-3.5, GPT-4, OpenAI, Claude, Anthropic, Gemini, Grok, Bard, PaLM, Llama, Copilot, “large language model”, LLM, “generative AI”, and “AI-assisted” writing. Potential matches were manually reviewed to confirm that surrounding context indicated use for writing or editing (for example, “we used”, “assisted by”, or acknowledgments of writing assistance), rather than unrelated mentions. To contextualize these findings, we also reviewed the publicly available Instructions for Authors of each included journal to identify policies related to disclosure of LLM-assisted writing. None of the included journals explicitly required disclosure of generative-AI or LLM-assisted writing in their posted author guidelines at the time of review. We selected language features that could reflect patterns of AI-assisted drafting without relying on ophthalmology-specific content. This allowed us to capture general stylistic tendencies rather than specialty terminology and to compare articles across journals within a common field-bounded corpus. Discourse and editorial markers that are frequently introduced or emphasized by automated rewriting tools were measured as document-level indicators, meaning features evaluated across the entire article rather than within individual sentences. These included connective adverbs such as “furthermore,” “moreover,” and “notably,” the formulaic phrase “in this study,” and editorial nouns such as “methodology.” To operationalize these signals systematically, we used a prespecified lexicon of 223 terms representing candidate discourse and editorial markers associated with AI-assisted academic writing. The initial candidate list was generated by asking ChatGPT which phrases it commonly produces when drafting scientific text. The resulting list was fixed before analysis to ensure reproducibility. Because these expressions are not unique to LLM-generated text, the lexicon was used as an indicator of stylistic prevalence rather than as a diagnostic classifier of authorship. Additional examples from published commentary on automated text generation and prior analyses of LLM-associated phrasing were reviewed to confirm that the ChatGPT-derived terms aligned with patterns described in the literature.(Amirjalili et al. 2024 ; Bao et al. 2025 ; Ji et al. 2023 ) In addition to evaluating individual terms, we summarized the overall frequency of lexicon terms as rates per 1,000 words at the journal level. We also examined 20 punctuation categories and calculated their frequency per 1,000 words, as punctuation patterns reflect sentence restructuring and revision practices commonly associated with automated editing. These categories represented the standard punctuation marks reliably identifiable during text extraction. A composite stylistic divergence index was constructed from punctuation rates and lexicon features. Within each journal quartile, individual features were standardized using pre-LLM means and standard deviations. These standardized values were then averaged to yield a document-level index, with higher values indicating greater stylistic divergence from the pre-LLM baseline. Authorship affiliation pattern was defined using the institutional countries of the first and last authors. An English-dominant affiliation was defined as a country in which English is an official language and widely used for scholarly communication (e.g., United States, United Kingdom, Canada, Australia, New Zealand, Ireland, Singapore). Four configurations were formed: both English-dominant, first-English-only, last-English-only, and neither English-dominant. For post-LLM comparisons, the composite stylistic divergence index was summarized overall and by quartile for each configuration, using the both English-dominant configuration as the reference group. The primary outcomes were changes in lexicon phrase prevalence, punctuation rates, and the composite stylistic divergence index; authorship affiliation pattern and disclosure analyses were prespecified secondary analyses. Institutional Review Board approval was not required because this study analyzed publicly available, published articles and did not involve human subjects or identifiable private information. Statistical Analysis Feature extraction was performed in Python 3.11. Statistical analyses were conducted in IBM SPSS Statistics for Windows, Version 30.0 (IBM Corp., Armonk, NY, USA), and plots were generated in R version 4.5.1 (R Foundation for Statistical Computing, Vienna, Austria). Document-level prevalence of lexicon terms was compared between pre- and post-LLM periods using two-sided tests for independent proportions. Both nominal significance (p < 0.05) and FDR-adjusted significance were examined, with multiplicity controlled by the Benjamini–Hochberg false discovery rate at q = 0.05; q-values are reported. Stylistic analyses included ordinary least squares regression models for the composite stylistic divergence index and for individual features, including 20 punctuation categories, sentence length, word length, and type-token ratio, with predictors for post-LLM period, quartile, and authorship affiliation pattern. Models were fit by quartile and pooled across quartiles; robust standard errors were used. Disclosure analyses were restricted to the post-LLM period. Disclosure prevalence was summarized descriptively by journal quartile, and logistic regression models were fit with disclosure (yes/no) as the outcome to assess differences across quartiles. Interrupted time-series regression was used to evaluate temporal trends in monthly disclosure prevalence after January 2023. For interrupted time-series analyses of the composite stylistic divergence index, monthly means were computed by quartile; pre-LLM months (January 2018–December 2020) were coded as zero to represent baseline, and intervening 2021–2022 months were not sampled. In a prespecified post-LLM cross-sectional contrast, composite means were compared across the four authorship-origin configurations (Overall and Q1–Q4) using the both English-dominant configuration as the reference, with two-sided independent-samples tests and unequal-variance (Welch) corrections. Two-sided α = 0.05 defined statistical significance; subgroups with n < 2 were not tested, and very small subgroups (n < 5) are flagged in results to denote reduced precision. Effect estimates and confidence intervals were reported to two decimal places, except for interrupted time-series slopes, which were reported to three decimals due to their small magnitude. Results Corpus characteristics The final dataset included 862 original research articles. By quartile, pre- and post-LLM counts were as follows: Q1, 143 and 124; Q2, 108 and 93; Q3, 108 and 93; and Q4, 105 and 88 (Table 1 ). Table 1 Corpus counts by journal and era. Journal Journal Quartile Pre-2023 (Baseline) Post-2023 (LLM Era) Ophthalmology Q1 143 124 Graefe’s Archive for Clinical and Experimental Ophthalmology Q2 108 93 BMC Ophthalmology Q3 108 93 Turkish Journal of Ophthalmology Q4 105 88 Counts represent the number of original research articles analyzed in each included journal (quartile shown). Pre-2023 = January 2018–December 2020; Post-2023 = January 2023–July 2025 (LLM era). Total corpus size across all quartiles and eras was N = 862 articles. Term-level prevalence shifts Phrase-level differences in prevalence were common across quartiles. At a nominal threshold of p < 0.05, 61 phrases showed significant differences in prevalence between pre- and post-LLM periods: 15 in Q1, 8 in Q2, 15 in Q3, and 23 in Q4. After controlling for multiple testing with a false discovery rate (FDR) of q < 0.05, 10 phrases remained significant. Q1 retained six phrases (four increased and two decreased), Q3 retained one increase, and Q4 retained three (two increases and one decrease). Q2 had no phrases that remained significant after FDR correction. Examples of notable changes include “in this study” in Q1, which increased from a prevalence of 0.92 pre-LLM to 1.00 post-LLM (q = 0.04), “prompt” in Q1 (0.03 vs. 0.14, q = 0.04), and “leveraging” in Q1 (0.00 vs. 0.07, q = 0.04). In Q3, the term “mitigate” rose from 0.00 to 0.12 (q = 0.03). In Q4, “notably” increased from 0.03 to 0.19 (q = 0.02) and “methodology” from 0.04 to 0.19 (q = 0.02). The distribution of log₂(post:pre) changes plotted against − log₁₀(q) values showed right-shifted tails in Q3 and Q4, consistent with stronger adoption of discourse and editorial markers in lower-tier journals (Fig. 1 ; Supplementary Table S1 ). Aggregate lexicon rates by quartile At the journal level, lexicon use summarized as rates per 1,000 words declined in Q1 and Q2 but increased in Q3 and Q4. In Q1, rates decreased from 5.88 to 5.12, with a post:pre rate ratio of 0.87 (95% CI: 0.84 to 0.91). In Q2, the decline was from 6.54 to 5.88, with a ratio of 0.90 (95% CI: 0.86 to 0.94). In Q3, rates increased from 6.01 to 6.50, with a ratio of 1.08 (95% CI: 1.03 to 1.14). In Q4, rates rose from 5.39 to 5.92, with a ratio of 1.10 (95% CI: 1.04 to 1.16). This pattern was consistent with stronger post-LLM increases in aggregate lexicon rates in Q3 and Q4 than in Q1 and Q2 (Table 2 ). Beyond lexical markers, structural features of writing also shifted. Table 2 Lexicon rates, differences, and rate ratios by quartile. Journal Quartile Pre rate Post rate Δ (Post–Pre) % change Rate ratio 95% CI Q1 5.88 5.12 -0.76 -0.13 0.87 (0.84 to 0.91) Q2 6.54 5.88 -0.66 -0.10 0.9 (0.86 to 0.94) Q3 6.01 6.5 + 0.49 + 0.08 1.08 (1.03 to 1.14) Q4 5.39 5.92 + 0.53 + 0.10 1.1 (1.04 to 1.16) Rates are expressed per 1,000 words (word tokens derived from whitespace-delimited tokenization of extracted body text). Pre refers to the pre-LLM period (2018–2020), and Post refers to the LLM era (2023–2025). Δ (Post–Pre) represents the absolute difference in rates. % change represents the proportional change from the pre- to post-period. Rate ratios are shown with 95% confidence intervals. Punctuation and structural style indicators Punctuation analyses revealed marked post-LLM changes in dash usage, particularly in Q4. For the em dash, post:pre rate ratios (95% CI) were 0.90 (0.35 to 2.33) in Q1, 0.65 (0.50 to 0.85) in Q2, 1.49 (0.88 to 2.51) in Q3, and 28.87 (1.73 to 483.30) in Q4. For the en dash, the corresponding ratios were 0.65 (0.54 to 0.79) in Q1, 0.89 (0.86 to 0.93) in Q2, 0.81 (0.78 to 0.85) in Q3, and 12.06 (9.25 to 15.72) in Q4. These large confidence intervals reflect the very low baseline frequency of dashes in pre-LLM articles, which explains the extreme ratios observed in Q4. Commas showed small but consistent increases in Q2 through Q4, while Q1 remained stable. Colons increased in Q2 but declined in Q1 and Q3. A forest plot illustrating the most pronounced punctuation changes is presented in Fig. 2 , with complete results available in Supplementary Table S2 . At the article level, regression models pooling across quartiles indicated that sentence length did not change significantly (95% CI: −0.44 to 0.30, p = 0.72). Average word length decreased slightly (95% CI: −0.07 to − 0.00, p = 0.03). The type–token ratio declined but did not reach statistical significance (95% CI: −0.01 to 0.00, p = 0.18). Full regression results are reported in Supplementary Table S3 . Disclosure of generative-AI use Analyses of explicit disclosure were limited to the post-LLM period. Overall, 10 of 398 articles (2.51%, 95% CI: 1.37 to 4.56) reported LLM use. Disclosure rates by quartile were 3.23% in Q1 (n = 124), 1.08% in Q2 (n = 93), 2.15% in Q3 (n = 93), and 3.41% in Q4 (n = 88) (Table 3 ). Logistic regression models showed no significant differences in disclosure across quartiles. For example, Q4 vs. Q1 yielded an odds ratio of 1.06 (95% CI: 0.23 to 4.85, p = 0.94), and Q4 vs. Q1–Q3 combined yielded 1.53 (95% CI: 0.39 to 6.03, p = 0.55). We next examined whether stylistic divergence varied across authorship affiliation patterns. Table 3 Explicit disclosures of LLM use in the post-LLM period, by quartile. Journal Quartile N No. % 95% CI Q1 124 4 3.23 (1.26 to 8.00) Q2 93 1 1.08 (0.19 to 5.84) Q3 93 2 2.15 (0.59 to 7.51) Q4 88 3 3.41 (1.17 to 9.55) Overall 398 10 2.51 (1.37 to 4.56) N = total number of post-LLM (2023–2025) articles in each quartile. No. = number of articles with explicit disclosure of LLM use. CI = confidence interval. Percentages represent the proportion of articles with explicit disclosure. Authorship affiliation pattern Analyses of authorship origin were also restricted to the post-LLM period. Articles with both first and last authors affiliated with English-dominant institutions, the reference group, had a mean composite stylistic divergence index of 0.11 standard deviations (95% CI: −0.01 to 0.24; n = 94). Articles with neither author from an English-dominant institution had a mean of 0.16 (95% CI: 0.12 to 0.19; n = 287), a difference that was not statistically significant (p = 0.53). Mixed configurations showed no consistent deviation from the reference: first-English-only articles averaged 0.03 (95% CI: −0.05 to 0.12; n = 8), and last-English-only articles averaged 0.33 (95% CI: −0.26 to 0.93; n = 7). In Q4, however, articles authored by English-dominant first and last authors had a higher mean (0.55, 95% CI: 0.26 to 0.83; n = 5) compared with the Neither group (0.22, 95% CI: 0.13 to 0.32; n = 80). The small sample size of the reference group in Q4 limits interpretation, but the contrast reached nominal significance (p = 0.04). Full results are summarized in Table 4 . Table 4 Authorship origin configurations in the post-LLM period, by quartile. Authorship configuration Subgroup Mean 95% CI No. Δ vs Both English-dominant p-value Both English-dominant institutions Overall (post‑LLM era) 0.11 (-0.01 to 0.24) 94 — — Q1 0.01 (-0.02 to 0.03) 77 — — Q2 0.78 (-0.47 to 2.02) 9 — — Q3 0.11 (-0.02 to 0.23) 3† — — Q4 0.55 (0.26 to 0.83) 5 — — First English-dominant only Overall (post‑LLM era) 0.03 (-0.05 to 0.12) 8 -0.08 0.32 Q1 0.00 (-0.07 to 0.08) 7 + 0.00 0.94 Last English-dominant only Overall (post‑LLM era) 0.33 (-0.26 to 0.93) 7 + 0.22 0.48 Q1 0.04 (-0.06 to 0.14) 4† + 0.03 0.53 Q4 1.02 (-1.18 to 3.23) 2† + 0.48 0.67 Neither English-dominant Overall (post‑LLM era) 0.16 (0.12 to 0.19) 287 + 0.04 0.53 Q1 0.03 (-0.02 to 0.07) 36 + 0.02 0.48 Q2 0.12 (0.08 to 0.17) 82 -0.65 0.31 Q3 0.18 (0.10 to 0.25) 89 + 0.07 0.35 Q4 0.22 (0.13 to 0.32) 80 -0.32 0.04 “Both English-dominant” serves as the reference group. Δ indicates the difference in mean composite stylistic score relative to the reference group. No. denotes the number of articles. p-values are shown for exploratory comparison with the reference group. † indicates subgroups with N < 5 and should be interpreted with caution. Post-LLM refers to articles published during the LLM era (2023–2025). Composite stylistic divergence index The composite stylistic divergence index increased in Q2 through Q4, while Q1 remained stable. In regression models, the estimated change in Q1 was 0.00 (95% CI: −0.03 to 0.02, p = 0.75). In Q2, the estimate was 0.14 (95% CI: 0.05 to 0.23, p < 0.01). In Q3, the estimate was 0.14 (95% CI: 0.08 to 0.21, p < 0.01). In Q4, the estimate was 0.19 (95% CI: 0.10 to 0.28, p < 0.01). These results indicate that the most pronounced post-LLM stylistic changes occurred in Q4 (Fig. 3 ). Temporal trend analysis Interrupted time-series analyses assess changes in trends before and after a defined time point. Using this approach, monthly averages of the composite stylistic divergence index from January 2023 to June 2025 showed positive slopes in Q1, Q2, and Q4, but no significant change in Q3. Estimated slope values (95% CI) were 0.002 (0.000 to 0.004) in Q1, 0.010 (0.001 to 0.018) in Q2, − 0.001 (− 0.007 to 0.004) in Q3, and 0.013 (0.001 to 0.024) in Q4. Inspection of half-year averages confirmed consistent increases in Q1, Q2, and Q4 over the study period, while Q3 remained stable (Fig. 4 ). Discussion We examined whether scientific writing in a field-bounded corpus of ophthalmology journals showed measurable stylistic changes in the post-LLM era. The strongest changes were observed in Q3 and Q4 journals, where lexical and structural features associated with generative writing tools increased after 2023. From a scientometric perspective, these findings suggest that technological changes in manuscript preparation may leave detectable traces in the published scientific record, even when direct evidence of tool use is unavailable. Analyses of individual phrases showed the same general pattern as the broader stylistic measures. Terms such as “in this study,” “notably,” and “methodology” increased in Q3 and Q4, whereas Q1 and Q2 showed little change in journal-level rate metrics. Although causality cannot be inferred, the distribution is consistent with broader uptake of standardized phrasing in the post-LLM era. These findings do not identify LLM use in individual manuscripts; rather, they indicate that corpus-level language patterns may shift as generative tools become integrated into writing workflows. Punctuation patterns showed a similar trend. En and em dashes were more frequent after 2023, particularly in Q4. Dashes often increase when sentences are reorganized or clauses are inserted during revision, a pattern compatible with automated rewriting as well as editorial preference. Other marks, including commas and colons, changed more modestly but in consistent directions. Small shifts in these features suggest alterations in clause boundaries or list formatting as text is rephrased or condensed. Although each change is subtle, together they indicate a broader restructuring of sentences in the post-LLM period. Authorship-origin contrasts were small and inconsistent across quartiles. In Q4 the reference group was very small, which limits interpretation. Overall, the findings suggest that author English-dominance alone is unlikely to explain the observed stylistic shifts. This does not prove LLM use, but it is compatible with broader, cross-regional adoption of standardized drafting practices in the post-LLM era. One possible explanation is that generative tools may act as linguistic standardizers. Such tools may nudge manuscripts toward common phrasing and structural templates. Although the analysis was conducted within ophthalmology journals, the specialty serves here primarily as a bounded empirical corpus for studying broader changes in scholarly communication. Its combination of structured article formats, stable publication practices, and internationally distributed authorship makes it a useful setting for examining how generative tools may influence the language of scientific publishing more generally. Despite growing attention to generative tools in research, explicit disclosure remained rare in our corpus. Fewer than 3% of post-LLM-period articles acknowledged LLM use, and disclosure rates did not differ significantly by quartile or authorship affiliation pattern. This aligns with prior reports describing inconsistent policies, variable journal guidance, and limited editorial capacity to detect AI-assisted writing.(Huang et al. 2025 ; Misra and Chandwar 2023 ; Resnik and Hosseini 2025 ; Stokel-Walker 2023 ) In response to these developments, some publishers have adopted proprietary AI-detection tools such as GPTZero, Turnitin AI Detection, and Originality.AI. However, these systems operate without transparency and vary widely in performance depending on domain, text length, and editing level.(Pudasaini et al. 2025 ) Their binary output offers limited interpretability and may not provide reliable support for editorial decisions. Our corpus-level approach avoids these pitfalls by emphasizing observable, longitudinal changes in style and structure over algorithmic classification. LLMs are also entering peer review. Pilot evaluations indicate that automated reports can mimic the structure of human reviews but often lack subject-matter depth.(Zhu et al. 2025 ) This raises accountability concerns when such tools are used without oversight. If both authors and reviewers rely on similar systems, stylistic convergence may increase across the publication process, reinforcing recurring phrasing and formatting patterns. A broader concern is cultural as well as editorial, because language models can carry over the rhetorical conventions and value systems embedded in their training data into scientific writing. As generative tools become integrated into scientific workflows, their outputs may shape rhetorical norms and editorial expectations. Chubb et al. have cautioned that heavy reliance on automated drafting may favor formulaic language, align manuscripts to metrics rather than substance, and blur individual author voice.(Chubb et al. 2022 ) These risks extend to peer review and editorial behavior. Commentaries have argued for stronger oversight of AI use in publishing and clearer disclosure frameworks.(Frangou et al. 2025 ; Grünebaum et al. 2025 ) If reviewers regularly encounter AI-generated phrasing, they may unconsciously adopt and reward these patterns, accelerating the standardization of academic style. Tao et al. have shown that LLMs embed Western cultural values, suggesting that even unintentional exposure may influence linguistic and argumentative preferences.(Tao et al. 2024 ) Changes in scientific writing style, while subtle, may affect how the scientific record is interpreted, evaluated, and synthesized. Journals may therefore wish to consider standardized disclosure policies for LLM use, and future meta-research could continue to monitor whether generative tools are associated with increasing linguistic standardization across disciplines. As LLMs become more integrated into both authorship and peer review, longitudinal tracking of language patterns may provide a useful complement to conventional bibliometric indicators. Several limitations should be considered. First, we cannot directly verify which articles involved LLM use, since our analysis is based on linguistic patterns rather than declared tool use. Second, editorial policies and peer review practices may have evolved during the study period in ways that could confound results. Third, while the journals included were consistent in scope and accessible for text extraction, they do not capture the full breadth of ophthalmology publishing. Moreover, only one representative journal per quartile was analyzed, which may limit generalizability within each tier. Fourth, the period between 2021 and 2022 was not sampled, as noted earlier, to avoid ambiguity during the transitional phase of LLM adoption. Finally, institutional affiliation was used as a coarse proxy for authorship language environment. Future work could extend this approach to other specialties. Qualitative review of manuscripts may also help capture subtler rhetorical shifts beyond lexical and punctuation features. In summary, we observed quantitative shifts in lexical and punctuation features in the post-LLM period, with the largest changes in lower-quartile journals. Although these patterns cannot establish LLM use in individual papers, they are consistent with broader changes in writing practice associated with the growing availability of automated drafting tools. Monitoring such stylistic indicators may help meta-researchers, editors, and publishers better understand how generative systems are reshaping scholarly communication. Declarations Compliance with Ethical Standards Funding No funding was received for conducting this study. Competing interests The authors declare that they have no competing interests. Ethics approval This study analyzed publicly available published articles and did not involve human subjects or identifiable private information. Institutional review board approval was not required. Research involving human participants and/or animals This study did not involve human participants or animals. Informed consent Informed consent was not required because the study analyzed publicly available published articles. Data availability Derived article-level feature data are available from the corresponding author upon reasonable request. Code availability Custom Python code used for text extraction and feature analysis is available from the corresponding author upon reasonable request. Author Contribution Tom Kornhauser : Conceptualization, methodology, formal analysis, investigation, writing - original draft. Tolossa Tufa Regassa : Interpretation, writing - review and editing. Morris E. Hartstein : Supervision, interpretation, writing - review and editing. Acknowledgement The authors thank Yogev Giladi for his contribution to the conceptualization of this study. References Amirjalili, F., Neysani, M., & Nikbakht, A. (2024). Exploring the boundaries of authorship: a comparative analysis of AI-generated text and human academic writing in English literature. Frontiers in Education , 9 . ttps://doi.org/10.3389/feduc.2024.1347421 Bao, T., Zhao, Y., Mao, J., & Zhang, C. (2025). Examining linguistic shifts in academic writing before and after the launch of ChatGPT: a study on preprint papers. Scientometrics , 130 (7), 3597–3627. ttps://doi.org/10.1007/S11192-025-05341-Y Chubb, J., Cowling, P., & Reed, D. (2022). Speeding up to keep up: exploring the use of AI in the research process. AI & Society , 37 (4), 1439–1457. ttps://doi.org/10.1007/s00146-021-01259-0 Desaire, H., Isom, M., & Hua, D. (2024). Almost Nobody Is Using ChatGPT to Write Academic Science Papers (Yet). Big Data and Cognitive Computing , 8 (10), 133. ttps://doi.org/10.3390/bdcc8100133 Frangou, S., Volpe, U., & Fiorillo, A. (2025). AI in scientific writing and publishing: A call for critical engagement. European Psychiatry , 68 (1), e98. ttps://doi.org/10.1192/j.eurpsy.2025.10061 Grünebaum, A., Dudenhausen, J., & Chervenak, F. A. (2025). The FAIR framework: ethical hybrid peer review. Journal of Perinatal Medicine , (0), 1–7. ttps://doi.org/10.1515/jpm-2025-0285 Guerin, M. B., Flynn, T. H., Brady, J., & O’Brien, C. J. (2009). Worldwide geographical distribution of ophthalmology publications. International Ophthalmology , 29 (6), 511–516. Huang, W., Liang, Y., Wei, X., & Du, Y. (2025). Ophthalmology Journals’ Guidelines on Generative Artificial Intelligence: A Comprehensive Analysis. American Journal of Ophthalmology , 271 , 445–454. ttps://doi.org/10.1016/j.ajo.2024.12.021 Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys , 55 (12), 1–38. ttps://doi.org/10.1145/3571730 Li, L.-J., & Ge, G.-C. (2009). Genre analysis: Structural and linguistic evolution of the English-medium medical research article (1985–2004). English for Specific Purposes , 28 (2), 93–104. Misra, D. P., & Chandwar, K. (2023). ChatGPT, artificial intelligence and scientific writing: What authors, peer reviewers and editors should know. Journal of the Royal College of Physicians of Edinburgh , 53 (2), 90–93. ttps://doi.org/10.1177/14782715231181023 Nguyen, A., Hong, Y., Dang, B., & Huang, X. (2024). Human-AI collaboration patterns in AI-assisted academic writing. Studies in Higher Education , 49 (5), 847–864. ttps://doi.org/10.1080/03075079.2024.2323593 Pudasaini, S., Miralles, L., Lillis, D., & Salvador, M. L. (2025). Benchmarking AI Text Detection: Assessing Detectors Against New Datasets, Evasion Tactics, and Enhanced LLMs. In Proceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect) (pp. 68–77). Abu Dhabi, UAE. Resnik, D. B., & Hosseini, M. (2025). The ethics of using artificial intelligence in scientific research: new guidance needed for a new tool. AI and Ethics , 5 (2), 1499–1521. ttps://doi.org/10.1007/s43681-024-00493-8 Stokel-Walker, C. (2023). ChatGPT listed as author on research papers: many scientists disapprove. Nature , 613 (7945), 620–621. ttps://doi.org/10.1038/d41586-023-00107-z Tao, Y., Viberg, O., Baker, R. S., & Kizilcec, R. F. (2024). Cultural bias and cultural alignment of large language models. PNAS Nexus , 3 (9), pgae346. ttps://doi.org/10.1093/pnasnexus/pgae346 Zhu, L., Lai, Y., Xie, J., Mou, W., Huang, L., Qi, C., et al. (2025). Evaluating the potential risks of employing large language models in peer review. Clinical and Translational Discovery , 5 (4), e70067. ttps://doi.org/10.1002/ctd2.70067 Additional Declarations No competing interests reported. Supplementary Files SupplementaryTableS1sm.docx SupplementaryTableS2sm.docx SupplementaryTableS3sm.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9082140","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":613498495,"identity":"eda47c5b-613e-4101-a10b-14f1560dcdc5","order_by":0,"name":"Tom Kornhauser","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvUlEQVRIiWNgGAWjYJACZiA24AexEgqIUM4D0yLZANJiQIoWgwMgLjFa7Nm7Ez8XVNgYG59fnfjhgQGDPL/YAQK28JzdLD3jTJqZ2Y23myWADjOcOTuBgBaJ3G3MvG2HbcxunN0A0pJgcJtYLcYzzm7+QZIWMwP+3m1E2nIG6BeeM2nGEjd4t1kkGEgQ9gt7e+/GzzwVNob9/Wc33/xRYSPPL01ACwJIgFVKEKscBPgPkKJ6FIyCUTAKRhIAACohP2ysKQ1rAAAAAElFTkSuQmCC","orcid":"","institution":"Assaf Harofeh Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Tom","middleName":"","lastName":"Kornhauser","suffix":""},{"id":613498496,"identity":"7a417701-4bd4-483a-af21-ab4be4b3978d","order_by":1,"name":"Tolossa Tufa Regassa","email":"","orcid":"","institution":"Jimma University","correspondingAuthor":false,"prefix":"","firstName":"Tolossa","middleName":"Tufa","lastName":"Regassa","suffix":""},{"id":613498498,"identity":"4068f7b7-88ff-409b-97f1-ead1f3847b0a","order_by":2,"name":"Morris E. Hartstein","email":"","orcid":"","institution":"Assaf Harofeh Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Morris","middleName":"E.","lastName":"Hartstein","suffix":""}],"badges":[],"createdAt":"2026-03-10 09:39:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9082140/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9082140/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105844790,"identity":"5c827d6f-ee56-4ab8-9586-9b503c9c2177","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":114723,"visible":true,"origin":"","legend":"\u003cp\u003eVolcano plots of lexicon phrase prevalence comparing pre-LLM (2018 to 2020) and post-LLM (2023 to 2025) periods, shown separately for journal quartiles Q1 to Q4. The x axis shows log₂ post to pre fold change and the y axis shows minus log₁₀ q. Phrases significant after false discovery rate correction (q \u0026lt; 0.05) are shown in black, with filled markers for increases and hollow markers for decreases. Grey points are not significant. The vertical grey line marks no change. All 223 lexicon phrases are shown for each quartile. Detailed prevalence values and q statistics are available in Supplementary Table S1.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/b95388f41fff586b270c6912.png"},{"id":105844792,"identity":"2880e328-e6fa-4891-afd1-7d7432a0fc41","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":79785,"visible":true,"origin":"","legend":"\u003cp\u003eForest plot of the five most significant punctuation changes within each journal quartile comparing post-LLM to pre-LLM periods. Points show post to pre rate ratios on a log scale with 95 percent confidence intervals. Shapes represent quartiles (Q1 circle, Q2 square, Q3 triangle, Q4 diamond). Grey fill with black outline denotes point estimates. Panels are displayed by quartile and rows are ordered by absolute effect size. The dashed vertical line marks a null ratio of 1.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/a191277567e371a371eea135.png"},{"id":105844794,"identity":"e0e9bf7f-de63-44a2-9b05-89b53b5497a0","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":29723,"visible":true,"origin":"","legend":"\u003cp\u003eEstimated post-LLM change in composite stylistic divergence index by journal quartile. Each point shows the quartile specific difference between post-LLM (2023 to 2025) and pre-LLM (2018 to 2020) periods based on ordinary least squares regression. Error bars show 95 percent confidence intervals. The composite score reflects standardized shifts in lexical and punctuation features. A value of 0 indicates no change. Q4 shows the largest increase.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/1393fb65f7d20bc7e4f31473.png"},{"id":105844795,"identity":"066a94af-f88f-49a7-a070-7e403229d29d","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":47620,"visible":true,"origin":"","legend":"\u003cp\u003eHalf year means of language features associated with LLM use by journal quartile during the post-LLM period from January 2023 through June 2025. Values represent averages within each half year.\u003c/p\u003e","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/873dc5977239da499d140067.png"},{"id":106401623,"identity":"88eb8633-c70e-4b40-95c3-05aec17a0218","added_by":"auto","created_at":"2026-04-08 09:08:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1124105,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/768548db-420a-43d5-86dd-0b3f468a06b9.pdf"},{"id":105844791,"identity":"0df0b8cd-4f22-4b0d-a03d-1794a2043f18","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":96275,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS1sm.docx","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/858b9d7a910ffd208890855f.docx"},{"id":105844796,"identity":"08bc555d-8548-48ee-ab58-fc787f44517d","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":37405,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS2sm.docx","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/ce6b6d2d36e095fe9c23fb34.docx"},{"id":105844797,"identity":"09c9b4cc-be0d-4272-8bd9-52acbc63e492","added_by":"auto","created_at":"2026-03-31 17:35:45","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":30414,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS3sm.docx","url":"https://assets-eu.researchsquare.com/files/rs-9082140/v1/27b56b54054a51023b515c11.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Large Language Models and Shifts in Scholarly Writing Style: A Cross-Journal Quantitative Analysis of Ophthalmology Research Articles","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLarge language models (LLMs), such as ChatGPT, Claude, Gemini, and Grok, are increasingly being used to assist with drafting, revising, and refining scientific manuscripts. While these tools may improve efficiency, they may also alter how research is phrased, structured, and rhetorically framed, even when their use is not explicitly disclosed.(Bao et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Desaire et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Nguyen et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) Because scholarly communication depends on shared conventions of tone, clarity, and presentation, the growing use of LLM-assisted writing may influence linguistic patterns within the published scientific record.\u003c/p\u003e \u003cp\u003eMost discussion of LLM use in research has focused on ethics, transparency, disclosure, and authorship.(Misra and Chandwar \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Resnik and Hosseini \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Stokel-Walker \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) However, fewer studies have examined whether the widespread availability of these tools is associated with measurable changes in the language of published articles themselves. Quantitative analysis of corpus-level linguistic features may provide an indirect way to study evolving writing practices, including potential shifts that occur without explicit disclosure.\u003c/p\u003e \u003cp\u003eA field-bounded publication corpus provides a useful setting for examining such changes. Journals within a single biomedical specialty often share similar article structures, editorial conventions, and genre expectations, while still differing in impact level, authorship composition, and publication culture. Ophthalmology offers a suitable case study because it combines structured research reporting with internationally distributed authorship and clearly stratified journal quartiles.(Guerin et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Li and Ge \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2009\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eBuilding on earlier analyses of linguistic shifts associated with generative language models,(Amirjalili et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Bao et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Ji et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) we analyzed full-length original research articles from four ophthalmology journals spanning quartiles Q1 through Q4. Rather than attempting to identify individual LLM-generated manuscripts, we focused on corpus-level changes in lexical, discourse, and punctuation features that may reflect evolving patterns of scientific writing. Our goal was to quantify how scholarly writing style may be changing in the post-LLM period and to assess whether these changes vary across journal quartiles, disclosure behavior, and authorship-affiliation patterns.\u003c/p\u003e"},{"header":"Data and Methods","content":"\u003cp\u003eWe assembled a longitudinal full-text publication corpus of original research articles from four general ophthalmology journals spanning journal quartiles Q1 through Q4. Quartiles were defined using Clarivate Journal Citation Reports (JCR) journal quartile assignments based on Journal Impact Factor category ranking (accessed March 5, 2026). The included journals were Ophthalmology (Q1), Graefe\u0026rsquo;s Archive for Clinical and Experimental Ophthalmology (Q2), BMC Ophthalmology (Q3), and the Turkish Journal of Ophthalmology (Q4). To reduce confounding by subspecialty scope and heterogeneous formatting, we restricted selection within each quartile to high-volume, broadly scoped general ophthalmology journals that publish primarily in English and had consistent monthly publication with stable PDF article structure across both study eras. Within each quartile, we selected a single representative journal meeting these criteria to enable systematic journal-month sampling and to prioritize internal comparability within a bounded disciplinary corpus over breadth.\u003c/p\u003e \u003cp\u003eThe pre-LLM period was defined as January 2018 through December 2020 and the post-LLM period as January 2023 through July 2025. This temporal split was designed to capture writing trends before and after the widespread availability of generative LLMs, while omitting the transitional period of 2021\u0026ndash;2022, during which LLM exposure was growing but adoption remained inconsistent. For each journal and calendar month, two to four eligible original research articles were systematically sampled. Reviews, editorials, letters, case reports and case series, brief communications, comments or perspectives, protocols, conference abstracts, non-English texts, and supplements or special issues were excluded.\u003c/p\u003e \u003cp\u003eWe used a custom automated workflow written in Python to extract text from PDFs, remove headers, footers, and reference sections, and compute linguistic features. This pipeline extracted article body text, removed recurrent boilerplate, truncated text at mechanically identifiable reference sections when present, and tokenized the remaining body text so features could be counted consistently across documents. The extractor used open-source utilities (pdfminer.six, PyPDF2), and article-level word counts were recorded for normalization so that all linguistic features could be expressed relative to article length rather than as raw counts.\u003c/p\u003e \u003cp\u003eThe workflow performed text extraction and feature counting only. It did not involve training or fine-tuning any machine-learning model on publisher content, and it did not submit article text to external LLMs. Analyses were performed on locally stored copies of articles that were accessed through institutional subscription access or open-access availability, and results are reported only in aggregate feature counts and rates rather than redistributing full text. No full-text content is reproduced in this manuscript or shared as part of the study outputs; only derived, non-reconstructive feature measures and aggregate results are reported.\u003c/p\u003e \u003cp\u003eTo assess explicit disclosure of generative-tool use, we searched the full text of articles from the post-LLM period for a prespecified list of model and tool terms and common variants. Search terms included ChatGPT, GPT, GPT-3.5, GPT-4, OpenAI, Claude, Anthropic, Gemini, Grok, Bard, PaLM, Llama, Copilot, \u0026ldquo;large language model\u0026rdquo;, LLM, \u0026ldquo;generative AI\u0026rdquo;, and \u0026ldquo;AI-assisted\u0026rdquo; writing. Potential matches were manually reviewed to confirm that surrounding context indicated use for writing or editing (for example, \u0026ldquo;we used\u0026rdquo;, \u0026ldquo;assisted by\u0026rdquo;, or acknowledgments of writing assistance), rather than unrelated mentions.\u003c/p\u003e \u003cp\u003eTo contextualize these findings, we also reviewed the publicly available Instructions for Authors of each included journal to identify policies related to disclosure of LLM-assisted writing. None of the included journals explicitly required disclosure of generative-AI or LLM-assisted writing in their posted author guidelines at the time of review.\u003c/p\u003e \u003cp\u003eWe selected language features that could reflect patterns of AI-assisted drafting without relying on ophthalmology-specific content. This allowed us to capture general stylistic tendencies rather than specialty terminology and to compare articles across journals within a common field-bounded corpus. Discourse and editorial markers that are frequently introduced or emphasized by automated rewriting tools were measured as document-level indicators, meaning features evaluated across the entire article rather than within individual sentences. These included connective adverbs such as \u0026ldquo;furthermore,\u0026rdquo; \u0026ldquo;moreover,\u0026rdquo; and \u0026ldquo;notably,\u0026rdquo; the formulaic phrase \u0026ldquo;in this study,\u0026rdquo; and editorial nouns such as \u0026ldquo;methodology.\u0026rdquo;\u003c/p\u003e \u003cp\u003eTo operationalize these signals systematically, we used a prespecified lexicon of 223 terms representing candidate discourse and editorial markers associated with AI-assisted academic writing. The initial candidate list was generated by asking ChatGPT which phrases it commonly produces when drafting scientific text. The resulting list was fixed before analysis to ensure reproducibility. Because these expressions are not unique to LLM-generated text, the lexicon was used as an indicator of stylistic prevalence rather than as a diagnostic classifier of authorship. Additional examples from published commentary on automated text generation and prior analyses of LLM-associated phrasing were reviewed to confirm that the ChatGPT-derived terms aligned with patterns described in the literature.(Amirjalili et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Bao et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Ji et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) In addition to evaluating individual terms, we summarized the overall frequency of lexicon terms as rates per 1,000 words at the journal level.\u003c/p\u003e \u003cp\u003eWe also examined 20 punctuation categories and calculated their frequency per 1,000 words, as punctuation patterns reflect sentence restructuring and revision practices commonly associated with automated editing. These categories represented the standard punctuation marks reliably identifiable during text extraction.\u003c/p\u003e \u003cp\u003eA composite stylistic divergence index was constructed from punctuation rates and lexicon features. Within each journal quartile, individual features were standardized using pre-LLM means and standard deviations. These standardized values were then averaged to yield a document-level index, with higher values indicating greater stylistic divergence from the pre-LLM baseline.\u003c/p\u003e \u003cp\u003eAuthorship affiliation pattern was defined using the institutional countries of the first and last authors. An English-dominant affiliation was defined as a country in which English is an official language and widely used for scholarly communication (e.g., United States, United Kingdom, Canada, Australia, New Zealand, Ireland, Singapore). Four configurations were formed: both English-dominant, first-English-only, last-English-only, and neither English-dominant. For post-LLM comparisons, the composite stylistic divergence index was summarized overall and by quartile for each configuration, using the both English-dominant configuration as the reference group.\u003c/p\u003e \u003cp\u003eThe primary outcomes were changes in lexicon phrase prevalence, punctuation rates, and the composite stylistic divergence index; authorship affiliation pattern and disclosure analyses were prespecified secondary analyses.\u003c/p\u003e \u003cp\u003eInstitutional Review Board approval was not required because this study analyzed publicly available, published articles and did not involve human subjects or identifiable private information.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eFeature extraction was performed in Python 3.11. Statistical analyses were conducted in IBM SPSS Statistics for Windows, Version 30.0 (IBM Corp., Armonk, NY, USA), and plots were generated in R version 4.5.1 (R Foundation for Statistical Computing, Vienna, Austria). Document-level prevalence of lexicon terms was compared between pre- and post-LLM periods using two-sided tests for independent proportions. Both nominal significance (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) and FDR-adjusted significance were examined, with multiplicity controlled by the Benjamini\u0026ndash;Hochberg false discovery rate at q\u0026thinsp;=\u0026thinsp;0.05; q-values are reported.\u003c/p\u003e \u003cp\u003eStylistic analyses included ordinary least squares regression models for the composite stylistic divergence index and for individual features, including 20 punctuation categories, sentence length, word length, and type-token ratio, with predictors for post-LLM period, quartile, and authorship affiliation pattern. Models were fit by quartile and pooled across quartiles; robust standard errors were used.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDisclosure\u003c/strong\u003e analyses were restricted to the post-LLM period. Disclosure prevalence was summarized descriptively by journal quartile, and logistic regression models were fit with disclosure (yes/no) as the outcome to assess differences across quartiles. Interrupted time-series regression was used to evaluate temporal trends in monthly disclosure prevalence after January 2023.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eFor interrupted time-series analyses of the composite stylistic divergence index, monthly means were computed by quartile; pre-LLM months (January 2018\u0026ndash;December 2020) were coded as zero to represent baseline, and intervening 2021\u0026ndash;2022 months were not sampled.\u003c/p\u003e \u003cp\u003eIn a prespecified post-LLM cross-sectional contrast, composite means were compared across the four authorship-origin configurations (Overall and Q1\u0026ndash;Q4) using the both English-dominant configuration as the reference, with two-sided independent-samples tests and unequal-variance (Welch) corrections. Two-sided α\u0026thinsp;=\u0026thinsp;0.05 defined statistical significance; subgroups with n\u0026thinsp;\u0026lt;\u0026thinsp;2 were not tested, and very small subgroups (n\u0026thinsp;\u0026lt;\u0026thinsp;5) are flagged in results to denote reduced precision. Effect estimates and confidence intervals were reported to two decimal places, except for interrupted time-series slopes, which were reported to three decimals due to their small magnitude.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eCorpus characteristics\u003c/h2\u003e \u003cp\u003eThe final dataset included 862 original research articles. By quartile, pre- and post-LLM counts were as follows: Q1, 143 and 124; Q2, 108 and 93; Q3, 108 and 93; and Q4, 105 and 88 (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCorpus counts by journal and era.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJournal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eJournal Quartile\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePre-2023 (Baseline)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePost-2023 (LLM Era)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOphthalmology\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e143\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e124\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGraefe\u0026rsquo;s Archive for Clinical and Experimental Ophthalmology\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e108\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBMC Ophthalmology\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e108\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTurkish Journal of Ophthalmology\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e105\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003eCounts represent the number of original research articles analyzed in each included journal (quartile shown). Pre-2023\u0026thinsp;=\u0026thinsp;January 2018\u0026ndash;December 2020; Post-2023\u0026thinsp;=\u0026thinsp;January 2023\u0026ndash;July 2025 (LLM era). \u003cb\u003eTotal corpus size across all quartiles and eras was N\u0026thinsp;=\u0026thinsp;862 articles.\u003c/b\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eTerm-level prevalence shifts\u003c/h3\u003e\n\u003cp\u003ePhrase-level differences in prevalence were common across quartiles. At a nominal threshold of p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, 61 phrases showed significant differences in prevalence between pre- and post-LLM periods: 15 in Q1, 8 in Q2, 15 in Q3, and 23 in Q4. After controlling for multiple testing with a false discovery rate (FDR) of q\u0026thinsp;\u0026lt;\u0026thinsp;0.05, 10 phrases remained significant. Q1 retained six phrases (four increased and two decreased), Q3 retained one increase, and Q4 retained three (two increases and one decrease). Q2 had no phrases that remained significant after FDR correction.\u003c/p\u003e \u003cp\u003eExamples of notable changes include \u0026ldquo;in this study\u0026rdquo; in Q1, which increased from a prevalence of 0.92 pre-LLM to 1.00 post-LLM (q\u0026thinsp;=\u0026thinsp;0.04), \u0026ldquo;prompt\u0026rdquo; in Q1 (0.03 vs. 0.14, q\u0026thinsp;=\u0026thinsp;0.04), and \u0026ldquo;leveraging\u0026rdquo; in Q1 (0.00 vs. 0.07, q\u0026thinsp;=\u0026thinsp;0.04). In Q3, the term \u0026ldquo;mitigate\u0026rdquo; rose from 0.00 to 0.12 (q\u0026thinsp;=\u0026thinsp;0.03). In Q4, \u0026ldquo;notably\u0026rdquo; increased from 0.03 to 0.19 (q\u0026thinsp;=\u0026thinsp;0.02) and \u0026ldquo;methodology\u0026rdquo; from 0.04 to 0.19 (q\u0026thinsp;=\u0026thinsp;0.02). The distribution of log₂(post:pre) changes plotted against \u0026minus;\u0026thinsp;log₁₀(q) values showed right-shifted tails in Q3 and Q4, consistent with stronger adoption of discourse and editorial markers in lower-tier journals (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e; Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eAggregate lexicon rates by quartile\u003c/h3\u003e\n\u003cp\u003eAt the journal level, lexicon use summarized as rates per 1,000 words declined in Q1 and Q2 but increased in Q3 and Q4. In Q1, rates decreased from 5.88 to 5.12, with a post:pre rate ratio of 0.87 (95% CI: 0.84 to 0.91). In Q2, the decline was from 6.54 to 5.88, with a ratio of 0.90 (95% CI: 0.86 to 0.94). In Q3, rates increased from 6.01 to 6.50, with a ratio of 1.08 (95% CI: 1.03 to 1.14). In Q4, rates rose from 5.39 to 5.92, with a ratio of 1.10 (95% CI: 1.04 to 1.16). This pattern was consistent with stronger post-LLM increases in aggregate lexicon rates in Q3 and Q4 than in Q1 and Q2 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Beyond lexical markers, structural features of writing also shifted.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eLexicon rates, differences, and rate ratios by quartile.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJournal Quartile\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePre rate\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePost rate\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eΔ (Post\u0026ndash;Pre)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e% change\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRate ratio\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e-0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e(0.84 to 0.91)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e-0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e(0.86 to 0.94)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e+\u0026thinsp;0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e(1.03 to 1.14)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e+\u0026thinsp;0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e(1.04 to 1.16)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003eRates are expressed per 1,000 words (word tokens derived from whitespace-delimited tokenization of extracted body text). Pre refers to the pre-LLM period (2018\u0026ndash;2020), and Post refers to the LLM era (2023\u0026ndash;2025). Δ (Post\u0026ndash;Pre) represents the absolute difference in rates. % change represents the proportional change from the pre- to post-period. Rate ratios are shown with 95% confidence intervals.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePunctuation and structural style indicators\u003c/h2\u003e \u003cp\u003ePunctuation analyses revealed marked post-LLM changes in dash usage, particularly in Q4. For the em dash, post:pre rate ratios (95% CI) were 0.90 (0.35 to 2.33) in Q1, 0.65 (0.50 to 0.85) in Q2, 1.49 (0.88 to 2.51) in Q3, and 28.87 (1.73 to 483.30) in Q4. For the en dash, the corresponding ratios were 0.65 (0.54 to 0.79) in Q1, 0.89 (0.86 to 0.93) in Q2, 0.81 (0.78 to 0.85) in Q3, and 12.06 (9.25 to 15.72) in Q4. These large confidence intervals reflect the very low baseline frequency of dashes in pre-LLM articles, which explains the extreme ratios observed in Q4.\u003c/p\u003e \u003cp\u003eCommas showed small but consistent increases in Q2 through Q4, while Q1 remained stable. Colons increased in Q2 but declined in Q1 and Q3. A forest plot illustrating the most pronounced punctuation changes is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, with complete results available in Supplementary Table \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAt the article level, regression models pooling across quartiles indicated that sentence length did not change significantly (95% CI: \u0026minus;0.44 to 0.30, p\u0026thinsp;=\u0026thinsp;0.72). Average word length decreased slightly (95% CI: \u0026minus;0.07 to \u0026minus;\u0026thinsp;0.00, p\u0026thinsp;=\u0026thinsp;0.03). The type\u0026ndash;token ratio declined but did not reach statistical significance (95% CI: \u0026minus;0.01 to 0.00, p\u0026thinsp;=\u0026thinsp;0.18). Full regression results are reported in Supplementary Table \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDisclosure of generative-AI use\u003c/h3\u003e\n\u003cp\u003eAnalyses of explicit disclosure were limited to the post-LLM period. Overall, 10 of 398 articles (2.51%, 95% CI: 1.37 to 4.56) reported LLM use. Disclosure rates by quartile were 3.23% in Q1 (n\u0026thinsp;=\u0026thinsp;124), 1.08% in Q2 (n\u0026thinsp;=\u0026thinsp;93), 2.15% in Q3 (n\u0026thinsp;=\u0026thinsp;93), and 3.41% in Q4 (n\u0026thinsp;=\u0026thinsp;88) (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Logistic regression models showed no significant differences in disclosure across quartiles. For example, Q4 vs. Q1 yielded an odds ratio of 1.06 (95% CI: 0.23 to 4.85, p\u0026thinsp;=\u0026thinsp;0.94), and Q4 vs. Q1\u0026ndash;Q3 combined yielded 1.53 (95% CI: 0.39 to 6.03, p\u0026thinsp;=\u0026thinsp;0.55). We next examined whether stylistic divergence varied across authorship affiliation patterns.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExplicit disclosures of LLM use in the post-LLM period, by quartile.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJournal Quartile\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e124\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e3.23\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(1.26 to 8.00)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1.08\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(0.19 to 5.84)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e2.15\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(0.59 to 7.51)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQ4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e3.41\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(1.17 to 9.55)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e398\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e2.51\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e(1.37 to 4.56)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003cb\u003eN\u003c/b\u003e\u0026thinsp;=\u0026thinsp;total number of post-LLM (2023\u0026ndash;2025) articles in each quartile. \u003cb\u003eNo.\u003c/b\u003e = number of articles with explicit disclosure of LLM use. \u003cb\u003eCI\u003c/b\u003e\u0026thinsp;=\u0026thinsp;confidence interval. Percentages represent the proportion of articles with explicit disclosure.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eAuthorship affiliation pattern\u003c/h3\u003e\n\u003cp\u003eAnalyses of authorship origin were also restricted to the post-LLM period. Articles with both first and last authors affiliated with English-dominant institutions, the reference group, had a mean composite stylistic divergence index of 0.11 standard deviations (95% CI: \u0026minus;0.01 to 0.24; n\u0026thinsp;=\u0026thinsp;94). Articles with neither author from an English-dominant institution had a mean of 0.16 (95% CI: 0.12 to 0.19; n\u0026thinsp;=\u0026thinsp;287), a difference that was not statistically significant (p\u0026thinsp;=\u0026thinsp;0.53). Mixed configurations showed no consistent deviation from the reference: first-English-only articles averaged 0.03 (95% CI: \u0026minus;0.05 to 0.12; n\u0026thinsp;=\u0026thinsp;8), and last-English-only articles averaged 0.33 (95% CI: \u0026minus;0.26 to 0.93; n\u0026thinsp;=\u0026thinsp;7).\u003c/p\u003e \u003cp\u003eIn Q4, however, articles authored by English-dominant first and last authors had a higher mean (0.55, 95% CI: 0.26 to 0.83; n\u0026thinsp;=\u0026thinsp;5) compared with the Neither group (0.22, 95% CI: 0.13 to 0.32; n\u0026thinsp;=\u0026thinsp;80). The small sample size of the reference group in Q4 limits interpretation, but the contrast reached nominal significance (p\u0026thinsp;=\u0026thinsp;0.04). Full results are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAuthorship origin configurations in the post-LLM period, by quartile.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAuthorship configuration\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubgroup\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMean\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eΔ vs Both English-dominant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003eBoth English-dominant institutions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall (post‑LLM era)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.01 to 0.24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.02 to 0.03)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.47 to 2.02)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.02 to 0.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(0.26 to 0.83)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eFirst English-dominant only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall (post‑LLM era)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.05 to 0.12)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.07 to 0.08)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eLast English-dominant only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall (post‑LLM era)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.26 to 0.93)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.06 to 0.14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-1.18 to 3.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e \u003cp\u003eNeither English-dominant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall (post‑LLM era)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(0.12 to 0.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e287\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(-0.02 to 0.07)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(0.08 to 0.17)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.31\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(0.10 to 0.25)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQ4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(0.13 to 0.32)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u0026ldquo;Both English-dominant\u0026rdquo; serves as the reference group. Δ indicates the difference in mean composite stylistic score relative to the reference group. \u003cb\u003eNo.\u003c/b\u003e denotes the number of articles. p-values are shown for exploratory comparison with the reference group. \u0026dagger; indicates subgroups with N\u0026thinsp;\u0026lt;\u0026thinsp;5 and should be interpreted with caution. \u003cem\u003ePost-LLM refers to articles published during the LLM era (2023\u0026ndash;2025).\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eComposite stylistic divergence index\u003c/h2\u003e \u003cp\u003eThe composite stylistic divergence index increased in Q2 through Q4, while Q1 remained stable. In regression models, the estimated change in Q1 was 0.00 (95% CI: \u0026minus;0.03 to 0.02, p\u0026thinsp;=\u0026thinsp;0.75). In Q2, the estimate was 0.14 (95% CI: 0.05 to 0.23, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). In Q3, the estimate was 0.14 (95% CI: 0.08 to 0.21, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). In Q4, the estimate was 0.19 (95% CI: 0.10 to 0.28, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). These results indicate that the most pronounced post-LLM stylistic changes occurred in Q4 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eTemporal trend analysis\u003c/h2\u003e \u003cp\u003eInterrupted time-series analyses assess changes in trends before and after a defined time point. Using this approach, monthly averages of the composite stylistic divergence index from January 2023 to June 2025 showed positive slopes in Q1, Q2, and Q4, but no significant change in Q3. Estimated slope values (95% CI) were 0.002 (0.000 to 0.004) in Q1, 0.010 (0.001 to 0.018) in Q2, \u0026minus;\u0026thinsp;0.001 (\u0026minus;\u0026thinsp;0.007 to 0.004) in Q3, and 0.013 (0.001 to 0.024) in Q4. Inspection of half-year averages confirmed consistent increases in Q1, Q2, and Q4 over the study period, while Q3 remained stable (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe examined whether scientific writing in a field-bounded corpus of ophthalmology journals showed measurable stylistic changes in the post-LLM era. The strongest changes were observed in Q3 and Q4 journals, where lexical and structural features associated with generative writing tools increased after 2023. From a scientometric perspective, these findings suggest that technological changes in manuscript preparation may leave detectable traces in the published scientific record, even when direct evidence of tool use is unavailable.\u003c/p\u003e \u003cp\u003eAnalyses of individual phrases showed the same general pattern as the broader stylistic measures. Terms such as \u0026ldquo;in this study,\u0026rdquo; \u0026ldquo;notably,\u0026rdquo; and \u0026ldquo;methodology\u0026rdquo; increased in Q3 and Q4, whereas Q1 and Q2 showed little change in journal-level rate metrics. Although causality cannot be inferred, the distribution is consistent with broader uptake of standardized phrasing in the post-LLM era. These findings do not identify LLM use in individual manuscripts; rather, they indicate that corpus-level language patterns may shift as generative tools become integrated into writing workflows.\u003c/p\u003e \u003cp\u003ePunctuation patterns showed a similar trend. En and em dashes were more frequent after 2023, particularly in Q4. Dashes often increase when sentences are reorganized or clauses are inserted during revision, a pattern compatible with automated rewriting as well as editorial preference. Other marks, including commas and colons, changed more modestly but in consistent directions. Small shifts in these features suggest alterations in clause boundaries or list formatting as text is rephrased or condensed. Although each change is subtle, together they indicate a broader restructuring of sentences in the post-LLM period.\u003c/p\u003e \u003cp\u003eAuthorship-origin contrasts were small and inconsistent across quartiles. In Q4 the reference group was very small, which limits interpretation. Overall, the findings suggest that author English-dominance alone is unlikely to explain the observed stylistic shifts. This does not prove LLM use, but it is compatible with broader, cross-regional adoption of standardized drafting practices in the post-LLM era. One possible explanation is that generative tools may act as linguistic standardizers. Such tools may nudge manuscripts toward common phrasing and structural templates.\u003c/p\u003e \u003cp\u003eAlthough the analysis was conducted within ophthalmology journals, the specialty serves here primarily as a bounded empirical corpus for studying broader changes in scholarly communication. Its combination of structured article formats, stable publication practices, and internationally distributed authorship makes it a useful setting for examining how generative tools may influence the language of scientific publishing more generally.\u003c/p\u003e \u003cp\u003eDespite growing attention to generative tools in research, explicit disclosure remained rare in our corpus. Fewer than 3% of post-LLM-period articles acknowledged LLM use, and disclosure rates did not differ significantly by quartile or authorship affiliation pattern. This aligns with prior reports describing inconsistent policies, variable journal guidance, and limited editorial capacity to detect AI-assisted writing.(Huang et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Misra and Chandwar \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Resnik and Hosseini \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Stokel-Walker \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2023\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eIn response to these developments, some publishers have adopted proprietary AI-detection tools such as GPTZero, Turnitin AI Detection, and Originality.AI. However, these systems operate without transparency and vary widely in performance depending on domain, text length, and editing level.(Pudasaini et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) Their binary output offers limited interpretability and may not provide reliable support for editorial decisions. Our corpus-level approach avoids these pitfalls by emphasizing observable, longitudinal changes in style and structure over algorithmic classification.\u003c/p\u003e \u003cp\u003eLLMs are also entering peer review. Pilot evaluations indicate that automated reports can mimic the structure of human reviews but often lack subject-matter depth.(Zhu et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) This raises accountability concerns when such tools are used without oversight. If both authors and reviewers rely on similar systems, stylistic convergence may increase across the publication process, reinforcing recurring phrasing and formatting patterns.\u003c/p\u003e \u003cp\u003eA broader concern is cultural as well as editorial, because language models can carry over the rhetorical conventions and value systems embedded in their training data into scientific writing. As generative tools become integrated into scientific workflows, their outputs may shape rhetorical norms and editorial expectations. Chubb et al. have cautioned that heavy reliance on automated drafting may favor formulaic language, align manuscripts to metrics rather than substance, and blur individual author voice.(Chubb et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) These risks extend to peer review and editorial behavior. Commentaries have argued for stronger oversight of AI use in publishing and clearer disclosure frameworks.(Frangou et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Gr\u0026uuml;nebaum et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) If reviewers regularly encounter AI-generated phrasing, they may unconsciously adopt and reward these patterns, accelerating the standardization of academic style. Tao et al. have shown that LLMs embed Western cultural values, suggesting that even unintentional exposure may influence linguistic and argumentative preferences.(Tao et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eChanges in scientific writing style, while subtle, may affect how the scientific record is interpreted, evaluated, and synthesized. Journals may therefore wish to consider standardized disclosure policies for LLM use, and future meta-research could continue to monitor whether generative tools are associated with increasing linguistic standardization across disciplines. As LLMs become more integrated into both authorship and peer review, longitudinal tracking of language patterns may provide a useful complement to conventional bibliometric indicators.\u003c/p\u003e \u003cp\u003eSeveral limitations should be considered. First, we cannot directly verify which articles involved LLM use, since our analysis is based on linguistic patterns rather than declared tool use. Second, editorial policies and peer review practices may have evolved during the study period in ways that could confound results. Third, while the journals included were consistent in scope and accessible for text extraction, they do not capture the full breadth of ophthalmology publishing. Moreover, only one representative journal per quartile was analyzed, which may limit generalizability within each tier. Fourth, the period between 2021 and 2022 was not sampled, as noted earlier, to avoid ambiguity during the transitional phase of LLM adoption. Finally, institutional affiliation was used as a coarse proxy for authorship language environment. Future work could extend this approach to other specialties. Qualitative review of manuscripts may also help capture subtler rhetorical shifts beyond lexical and punctuation features.\u003c/p\u003e \u003cp\u003eIn summary, we observed quantitative shifts in lexical and punctuation features in the post-LLM period, with the largest changes in lower-quartile journals. Although these patterns cannot establish LLM use in individual papers, they are consistent with broader changes in writing practice associated with the growing availability of automated drafting tools. Monitoring such stylistic indicators may help meta-researchers, editors, and publishers better understand how generative systems are reshaping scholarly communication.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003e\u003cu\u003eCompliance with Ethical Standards\u003c/u\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo funding was received for conducting this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study analyzed publicly available published articles and did not involve human subjects or identifiable private information. Institutional review board approval was not required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResearch involving human participants and/or animals\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study did not involve human participants or animals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed consent\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInformed consent was not required because the study analyzed publicly available published articles.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDerived article-level feature data are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCustom Python code used for text extraction and feature analysis is available from the corresponding author upon reasonable request.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eTom Kornhauser : Conceptualization, methodology, formal analysis, investigation, writing - original draft. Tolossa Tufa Regassa : Interpretation, writing - review and editing. Morris E. Hartstein : Supervision, interpretation, writing - review and editing.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors thank Yogev Giladi for his contribution to the conceptualization of this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAmirjalili, F., Neysani, M., \u0026amp; Nikbakht, A. (2024). Exploring the boundaries of authorship: a comparative analysis of AI-generated text and human academic writing in English literature. \u003cem\u003eFrontiers in Education\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e. ttps://doi.org/10.3389/feduc.2024.1347421\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBao, T., Zhao, Y., Mao, J., \u0026amp; Zhang, C. (2025). Examining linguistic shifts in academic writing before and after the launch of ChatGPT: a study on preprint papers. \u003cem\u003eScientometrics\u003c/em\u003e, \u003cem\u003e130\u003c/em\u003e(7), 3597\u0026ndash;3627. ttps://doi.org/10.1007/S11192-025-05341-Y\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChubb, J., Cowling, P., \u0026amp; Reed, D. (2022). Speeding up to keep up: exploring the use of AI in the research process. \u003cem\u003eAI \u0026amp; Society\u003c/em\u003e, \u003cem\u003e37\u003c/em\u003e(4), 1439\u0026ndash;1457. ttps://doi.org/10.1007/s00146-021-01259-0\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDesaire, H., Isom, M., \u0026amp; Hua, D. (2024). Almost Nobody Is Using ChatGPT to Write Academic Science Papers (Yet). \u003cem\u003eBig Data and Cognitive Computing\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(10), 133. ttps://doi.org/10.3390/bdcc8100133\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrangou, S., Volpe, U., \u0026amp; Fiorillo, A. (2025). AI in scientific writing and publishing: A call for critical engagement. \u003cem\u003eEuropean Psychiatry\u003c/em\u003e, \u003cem\u003e68\u003c/em\u003e(1), e98. ttps://doi.org/10.1192/j.eurpsy.2025.10061\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGr\u0026uuml;nebaum, A., Dudenhausen, J., \u0026amp; Chervenak, F. A. (2025). The FAIR framework: ethical hybrid peer review. \u003cem\u003eJournal of Perinatal Medicine\u003c/em\u003e, (0), 1\u0026ndash;7. ttps://doi.org/10.1515/jpm-2025-0285\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuerin, M. B., Flynn, T. H., Brady, J., \u0026amp; O\u0026rsquo;Brien, C. J. (2009). Worldwide geographical distribution of ophthalmology publications. \u003cem\u003eInternational Ophthalmology\u003c/em\u003e, \u003cem\u003e29\u003c/em\u003e(6), 511\u0026ndash;516.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, W., Liang, Y., Wei, X., \u0026amp; Du, Y. (2025). Ophthalmology Journals\u0026rsquo; Guidelines on Generative Artificial Intelligence: A Comprehensive Analysis. \u003cem\u003eAmerican Journal of Ophthalmology\u003c/em\u003e, \u003cem\u003e271\u003c/em\u003e, 445\u0026ndash;454. ttps://doi.org/10.1016/j.ajo.2024.12.021\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJi, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., et al. (2023). Survey of Hallucination in Natural Language Generation. \u003cem\u003eACM Computing Surveys\u003c/em\u003e, \u003cem\u003e55\u003c/em\u003e(12), 1\u0026ndash;38. ttps://doi.org/10.1145/3571730\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, L.-J., \u0026amp; Ge, G.-C. (2009). Genre analysis: Structural and linguistic evolution of the English-medium medical research article (1985\u0026ndash;2004). \u003cem\u003eEnglish for Specific Purposes\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(2), 93\u0026ndash;104.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMisra, D. P., \u0026amp; Chandwar, K. (2023). ChatGPT, artificial intelligence and scientific writing: What authors, peer reviewers and editors should know. \u003cem\u003eJournal of the Royal College of Physicians of Edinburgh\u003c/em\u003e, \u003cem\u003e53\u003c/em\u003e(2), 90\u0026ndash;93. ttps://doi.org/10.1177/14782715231181023\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen, A., Hong, Y., Dang, B., \u0026amp; Huang, X. (2024). Human-AI collaboration patterns in AI-assisted academic writing. \u003cem\u003eStudies in Higher Education\u003c/em\u003e, \u003cem\u003e49\u003c/em\u003e(5), 847\u0026ndash;864. ttps://doi.org/10.1080/03075079.2024.2323593\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePudasaini, S., Miralles, L., Lillis, D., \u0026amp; Salvador, M. L. (2025). Benchmarking AI Text Detection: Assessing Detectors Against New Datasets, Evasion Tactics, and Enhanced LLMs. In \u003cem\u003eProceedings of the 1st Workshop on GenAI Content Detection (GenAIDetect)\u003c/em\u003e (pp. 68\u0026ndash;77). Abu Dhabi, UAE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eResnik, D. B., \u0026amp; Hosseini, M. (2025). The ethics of using artificial intelligence in scientific research: new guidance needed for a new tool. \u003cem\u003eAI and Ethics\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(2), 1499\u0026ndash;1521. ttps://doi.org/10.1007/s43681-024-00493-8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStokel-Walker, C. (2023). ChatGPT listed as author on research papers: many scientists disapprove. \u003cem\u003eNature\u003c/em\u003e, \u003cem\u003e613\u003c/em\u003e(7945), 620\u0026ndash;621. ttps://doi.org/10.1038/d41586-023-00107-z\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTao, Y., Viberg, O., Baker, R. S., \u0026amp; Kizilcec, R. F. (2024). Cultural bias and cultural alignment of large language models. \u003cem\u003ePNAS Nexus\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(9), pgae346. ttps://doi.org/10.1093/pnasnexus/pgae346\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu, L., Lai, Y., Xie, J., Mou, W., Huang, L., Qi, C., et al. (2025). Evaluating the potential risks of employing large language models in peer review. \u003cem\u003eClinical and Translational Discovery\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(4), e70067. ttps://doi.org/10.1002/ctd2.70067\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Large language models, scholarly communication, scientometrics, text mining, scientific writing, bibliometrics","lastPublishedDoi":"10.21203/rs.3.rs-9082140/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9082140/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge language models (LLMs) are increasingly integrated into scientific writing workflows, raising questions about whether their widespread availability may influence the language of the published scientific record. We conducted a longitudinal text analysis to examine whether stylistic features of research articles changed following the introduction of widely accessible LLM tools. A corpus of 862 full-length original research articles was assembled from four general ophthalmology journals representing Clarivate Journal Citation Reports quartiles Q1\u0026ndash;Q4. Articles were sampled systematically by journal-month from pre-LLM (January 2018\u0026ndash;December 2020) and post-LLM (January 2023\u0026ndash;July 2025) periods. Using an automated text-processing workflow, we quantified lexical discourse markers and punctuation features associated with editorial and connective phrasing patterns in scientific writing. Feature frequencies were normalized by article length, and a composite stylistic divergence index was constructed using standardized feature values within each quartile. Post-LLM articles showed measurable stylistic shifts, most pronounced in Q3 and Q4 journals. Several discourse and editorial markers increased in prevalence, punctuation patterns shifted, and the composite stylistic divergence index increased significantly in lower-quartile journals while remaining stable in Q1. Explicit disclosure of generative tool use was rare, occurring in fewer than 3% of post-LLM articles. These findings suggest that corpus-level stylistic patterns in scientific writing may be evolving in the post-LLM era and illustrate how quantitative analysis of linguistic features can help monitor technological influences on scholarly communication.\u003c/p\u003e","manuscriptTitle":"Large Language Models and Shifts in Scholarly Writing Style: A Cross-Journal Quantitative Analysis of Ophthalmology Research Articles","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-31 17:35:41","doi":"10.21203/rs.3.rs-9082140/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2c7dc2ec-669c-480c-aace-4eb1fd30e7dc","owner":[],"postedDate":"March 31st, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-17T05:42:46+00:00","index":16,"fulltext":""},{"type":"reviewerAgreed","content":"167688339821768415426472692580917938212","date":"2026-05-11T23:54:34+00:00","index":13,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-31T17:35:41+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-31 17:35:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9082140","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9082140","identity":"rs-9082140","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}