Ophthalmology science · 2026
Evaluator-Masked Assessment of 4 Consumer-Facing Large Language Models on Expert Anterior Segment Ophthalmology Questions.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Yang Soonwon, Han Hyoji, Lee Hyun Soo
Аннотация
OBJECTIVE: To compare expert-rated quality and safety of responses from four contemporary large language models (LLMs)-ChatGPT, Gemini, Claude, and Grok-on expert-level anterior segment ophthalmology questions.
DESIGN: Cross-sectional, masked, randomized-display-order, multievaluator comparative study.
PARTICIPANTS: Eleven fellowship-trained cornea and anterior segment subspecialists.
METHODS: Forty-five expert-level questions spanning six anterior segment domains were administered to each LLM under zero-shot conditions in a single prespecified session (January 19, 2026), with memory, browsing, and retrieval disabled. Responses were masked to model identity and independently rated on five 5-point Likert metrics (expert-rated accuracy, completeness/clinical utility, clarity/readability, overall safety, and potential for harm). No specialist-authored reference answer key was used; all metrics index expert-rated perception rather than verified clinical correctness.
MAIN OUTCOME MEASURES: Median [interquartile range], mean ± standard deviation, favorable-rating proportion (≤2) with 95% bootstrap confidence interval, Friedman omnibus with Kendall's W, and Bonferroni-adjusted pairwise Wilcoxon tests with rank-biserial correlation. Prespecified sensitivity analyses included length-adjusted mixed models and a masking-integrity diagnostic; post hoc analyses added question-level severe-safety and style-restricted comparisons.
RESULTS: All 9900 ratings (45 questions × 4 LLMs × 5 metrics × 11 evaluators) were complete. Differences were significant for all three quality metrics (Friedman P < 0.001; Kendall's W = 0.44-0.54); Claude and Grok were each rated more favorably than ChatGPT and Gemini (all adjusted P < 0.001; |r_rb| = 0.73-0.98). No statistically significant between-model difference was detected in mean safety scores. Post hoc, more questions elicited ≥ 1 severe overall-safety rating for Claude (35.6%) and Grok (33.3%) than for ChatGPT (6.7%) or Gemini (11.1%) (Cochran Q = 23.4, P < 0.001), persisting when two-evaluator concurrence was required. Ordinal Krippendorff's α was 0.021-0.168; the 11-evaluator mean used as the primary analytic unit showed moderate reliability (intraclass correlation coefficient [ICC(2,k)] = 0.58-0.69 for quality metrics).
CONCLUSIONS: Claude and Grok received more favorable expert-rated quality than ChatGPT and Gemini. Post hoc, a larger severe-safety-rating tail was identified for these two models despite no significant difference in mean safety; this signal is hypothesis-generating. The findings reflect expert preference, not validated clinical correctness, at one time point.
FINANCIAL DISCLOSURES: The authors have no proprietary or commercial interest in any materials discussed in this article.
DESIGN: Cross-sectional, masked, randomized-display-order, multievaluator comparative study.
PARTICIPANTS: Eleven fellowship-trained cornea and anterior segment subspecialists.
METHODS: Forty-five expert-level questions spanning six anterior segment domains were administered to each LLM under zero-shot conditions in a single prespecified session (January 19, 2026), with memory, browsing, and retrieval disabled. Responses were masked to model identity and independently rated on five 5-point Likert metrics (expert-rated accuracy, completeness/clinical utility, clarity/readability, overall safety, and potential for harm). No specialist-authored reference answer key was used; all metrics index expert-rated perception rather than verified clinical correctness.
MAIN OUTCOME MEASURES: Median [interquartile range], mean ± standard deviation, favorable-rating proportion (≤2) with 95% bootstrap confidence interval, Friedman omnibus with Kendall's W, and Bonferroni-adjusted pairwise Wilcoxon tests with rank-biserial correlation. Prespecified sensitivity analyses included length-adjusted mixed models and a masking-integrity diagnostic; post hoc analyses added question-level severe-safety and style-restricted comparisons.
RESULTS: All 9900 ratings (45 questions × 4 LLMs × 5 metrics × 11 evaluators) were complete. Differences were significant for all three quality metrics (Friedman P < 0.001; Kendall's W = 0.44-0.54); Claude and Grok were each rated more favorably than ChatGPT and Gemini (all adjusted P < 0.001; |r_rb| = 0.73-0.98). No statistically significant between-model difference was detected in mean safety scores. Post hoc, more questions elicited ≥ 1 severe overall-safety rating for Claude (35.6%) and Grok (33.3%) than for ChatGPT (6.7%) or Gemini (11.1%) (Cochran Q = 23.4, P < 0.001), persisting when two-evaluator concurrence was required. Ordinal Krippendorff's α was 0.021-0.168; the 11-evaluator mean used as the primary analytic unit showed moderate reliability (intraclass correlation coefficient [ICC(2,k)] = 0.58-0.69 for quality metrics).
CONCLUSIONS: Claude and Grok received more favorable expert-rated quality than ChatGPT and Gemini. Post hoc, a larger severe-safety-rating tail was identified for these two models despite no significant difference in mean safety; this signal is hypothesis-generating. The findings reflect expert preference, not validated clinical correctness, at one time point.
FINANCIAL DISCLOSURES: The authors have no proprietary or commercial interest in any materials discussed in this article.