Examining ChatGPT Performance on USMLE Sample Items and Implications for Assessment

• Published in: Academic medicine Vol. 99; no. 2; pp. 192 - 197
• Main Authors: Yaneva, Victoria, Baldwin, Peter, Jurich, Daniel P, Swygert, Kimberly, Clauser, Brian E
• Format: Journal Article
• Language: English
• Published: United States Lippincott Williams & Wilkins 01-02-2024
• Subjects: Artificial Intelligence; Computer Simulation; Humans; Knowledge; Language; Learning; Research Reports
• ISSN: 1040-2446; 1938-808X; 1938-808X
• DOI: 10.1097/ACM.0000000000005549

In late 2022 and early 2023, reports that ChatGPT could pass the United States Medical Licensing Examination (USMLE) generated considerable excitement, and media response suggested ChatGPT has credible medical knowledge. This report analyzes the extent to which an artificial intelligence (AI) agent's performance on these sample items can generalize to performance on an actual USMLE examination and an illustration is given using ChatGPT.
As with earlier investigations, analyses were based on publicly available USMLE sample items. Each item was submitted to ChatGPT (version 3.5) 3 times to evaluate stability. Responses were scored following rules that match operational practice, and a preliminary analysis explored the characteristics of items that ChatGPT answered correctly. The study was conducted between February and March 2023.
For the full sample of items, ChatGPT scored above 60% correct except for one replication for Step 3. Response success varied across replications for 76 items (20%). There was a modest correspondence with item difficulty wherein ChatGPT was more likely to respond correctly to items found easier by examinees. ChatGPT performed significantly worse ( P < .001) on items relating to practice-based learning.
Achieving 60% accuracy is an approximate indicator of meeting the passing standard, requiring statistical adjustments for comparison. Hence, this assessment can only suggest consistency with the passing standards for Steps 1 and 2 Clinical Knowledge, with further limitations in extrapolating this inference to Step 3. These limitations are due to variances in item difficulty and exclusion of the simulation component of Step 3 from the evaluation-limitations that would apply to any AI system evaluated on the Step 3 sample items. It is crucial to note that responses from large language models exhibit notable variations when faced with repeated inquiries, underscoring the need for expert validation to ensure their utility as a learning tool.