Humans vs. large language models in neurology board examination: performance, limitations, and reference reliability

To evaluate the performance and reference reliability of three large language models in neurology using a national board examination framework.

A total of 803 validated multiple-choice questions from Turkish National Neurology Board Examinations (2015-2024) were administered to ChatGPT Plus, Gemini Advanced, and Microsoft Copilot Pro using a standardized prompt requiring an answer and a supporting reference. Model performance was compared with overall examinee performance and analyzed by neurological subspecialty, question type, and presence of visual content. References provided for incorrectly answered questions were independently evaluated by three board certified neurologists.

Mean accuracy rates were 87% for Gemini, 86% for Copilot, and 85% for ChatGPT, significantly outperforming the human examinee average of 65% (p < 0.001), with no significant differences among models. Accuracy did not differ by question type or neurological subspecialty. All models outperformed examinees on non-visual questions, whereas no performance advantage was observed for visually based items. Reference evaluation revealed substantial limitations: ChatGPT frequently provided insufficient citations (39.6%), while fabricated references predominated in Gemini (53.0%) and Co-pilot (42.1%).

Large language models demonstrate high and consistent accuracy on neurology board examination questions, with performance exceeding that of the average examinee. On visually based questions, accuracy was lower than for non-visual items, and the performance advantage over examinees disappeared. High rates of insufficient referencing indicate a clear need for expert oversight, supporting the use of LLMs as complementary tools in neurology education rather than autonomous sources of clinical or academic authority.