Re: Chen JL et al: Assessment of Correctness, Content Omission, and Risk of Harm in Large Language Model Responses to Ophthalmology Continuing Medical Education Questions Authors: James Vassallo Published: 2026 TO THE EDITOR: I read with great interest the study by Chen et al1 evaluating the performance of 2 old large language models (LLMs) in answering ophthalmology continuing medical education questions. However, this study is an example of how the traditional scientific publication process is unable to keep up with very rapid technological developments. This can lead readers to draw conclusions based on outdated studies. The study design by Chen et al goes beyond simple multiple-choice selection to investigate the qualitative aspects of artificial intelligence (AI) outputs. It demonstrates a significant performance gap between the models, with Google’s Gemini 1.5 Pro achieving a 49.2% accuracy rate, accompanied by a concerning 44% risk of introducing physical or emotional harm. The poor performance was particularly pronounced in the “Diagnosis” (36.0% accuracy) and “Pathophysiology” (42.3% accuracy) cohorts, which demand high-level reasoning. These metrics are a reflection of next-token prediction used by the models in this study. However, the rapid reiterations of these and other models mean that their performance would have already improved by the time of the manuscript's publication.2, 3, 4 Large language models have shifted from sequential token synthesis toward multitoken prediction models that execute parallel chain-of-thought processes before generating output. Hence, current models are capable of high-level reasoning and accessing current information. Studies published in 2026 utilizing the StatPearls ophthalmic question bank demonstrate improved performance in this domain. For example, Zhang et al2 showed an 83.3% diagnostic accuracy rate with Gemini 3 Flash. To bridge the lag between technological development and scientific publication, the ophthalmic community must pivot from static journal-published data sets toward dynamic, open-access, live benchmarking repositories that automatically test new models against validated clinical data sets upon release. Furthermore, the conversation must evolve from measuring whether an AI model can pass a board examination to formalizing the governance of its integration into day-to-day clinical practice, so that AI capabilities can be exploited on a larger scale for the benefit of our patients. Footnotes Disclosure(s): The author has completed and submitted the ICMJE disclosures form. The author has no proprietary or commercial interest in any materials discussed in this article. Declaration of Generative AI and AI-Assisted Technologies in the Writing Process. During the preparation of this work, the author used Google Gemini to get feedback on the manuscript draft. After using this tool/service, the author reviewed and edited the content as needed and takes full responsibility for the content of the publication. References