Abstract

Abstract Introduction: Large language models have demonstrated impressive capabilities, but application to medicine remains unclear. We seek to evaluate the use of ChatGPT on the American Urological Association Self-assessment Study Program as an educational adjunct for urology trainees and practicing physicians. Methods: One hundred fifty questions from the 2022 Self-assessment Study Program exam were screened, and those containing visual assets (n=15) were removed. The remaining items were encoded as open ended or multiple choice. ChatGPT’s output was coded as correct, incorrect, or indeterminate; if indeterminate, responses were regenerated up to 2 times. Concordance, quality, and accuracy were ascertained by 3 independent researchers and reviewed by 2 physician adjudicators. A new session was started for each entry to avoid crossover learning. Results: ChatGPT was correct on 36/135 (26.7%) open-ended and 38/135 (28.2%) multiple-choice questions. Indeterminate responses were generated in 40 (29.6%) and 4 (3.0%), respectively. Of the correct responses, 24/36 (66.7%) and 36/38 (94.7%) were on initial output, 8 (22.2%) and 1 (2.6%) on second output, and 4 (11.1%) and 1 (2.6%) on final output, respectively. Although regeneration decreased indeterminate responses, proportion of correct responses did not increase. For open-ended and multiple-choice questions, ChatGPT provided consistent justifications for incorrect answers and remained concordant between correct and incorrect answers. Conclusions: ChatGPT previously demonstrated promise on medical licensing exams; however, application to the 2022 Self-assessment Study Program was not demonstrated. Performance improved with multiple-choice over open-ended questions. More importantly were the persistent justifications for incorrect responses—left unchecked, utilization of ChatGPT in medicine may facilitate medical misinformation.