Lexical Diffusion: Linguistic Characteristics of Ophthalmology Science Articles with Disclosed Artificial...

To characterize discussion-section language in ophthalmology research articles with disclosed generative artificial intelligence (AI) writing assistance and to examine temporal lexical patterns among articles without disclosure.

Cross-sectional observational study.

Original research articles published in Ophthalmology Science from January 2024 through May 2026.

Articles were classified according to disclosure of generative AI writing assistance. Discussion-section text was analyzed for Flesch Reading Ease (FRE), Flesch-Kincaid Grade Level (FKGL), Simple Measure of Gobbledygook (SMOG), predefined generative AI-associated terms, control academic terms, and their ratio. Articles with disclosed AI writing assistance were compared with nondisclosing articles using repeated stratified 1:5 resampling by publication year and first-author English-language-dominant affiliation status across 1000 iterations. Temporal trends were evaluated among nondisclosing articles.

Flesch Reading Ease, FKGL, SMOG, term densities, and AI-associated-to-control term ratio.

Among 633 articles, 28 (4.4%) disclosed generative AI writing assistance. English-language-dominant affiliation fractions were lower among AI-disclosed articles for first and corresponding authors (P = 0.008 and 0.012, respectively). Artificial intelligence-disclosed articles had lower FRE (11.97 vs. 21.59), higher FKGL (17.15 vs. 16.00), and higher SMOG scores (17.82 vs. 17.12; P values < 0.001, 0.013, and 0.027, respectively). Resampling produced mean Hodges-Lehmann shifts of -7.35 for FRE (95% resampling interval, -8.90 to -5.83), 1.06 for FKGL (0.80-1.32), and 0.66 for SMOG (0.47-0.86). Artificial intelligence-associated term density and the AI-associated-to-control term ratio were higher in AI-disclosed articles (22.38 vs. 15.71 terms per 1000 words and 0.91 vs. 0.63; P = 0.003 and 0.004, respectively), whereas control-term density did not differ (P = 0.844); resampling confirmed these lexical findings. Among nondisclosing articles, AI-associated term density and the ratio increased over publication years (all, P < 0.001), whereas control-term density did not (all, P > 0.05).

Disclosed generative AI writing assistance was uncommon but was associated with lower formula-based readability, higher estimated reading grade levels, and a higher density of predefined generative AI-associated vocabulary in discussion sections. The temporal pattern among nondisclosing articles also suggests possible corpus-level diffusion of AI-associated lexical patterns, although disclosure status does not verify actual AI use.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.