Pure Wisdom or Potemkin Villages? A Comparison of ChatGPT 3.5 and ChatGPT 4 on USMLE Step 3 Style Questions: Quantitative Analysis
Open accessThe DOI matches its Crossref record; 6 of the 6 references it cites were found in a registry or an index.
Abstract
Background: The United States Medical Licensing Examination (USMLE) has been critical in medical education since 1992, testing various aspects of a medical student’s knowledge and skills through different steps, based on their training level. Artificial intelligence (AI) tools, including chatbots like ChatGPT, are emerging technologies with potential applications in medicine. However, comprehensive studies analyzing ChatGPT’s performance on USMLE Step 3 in large-scale scenarios and comparing different versions of ChatGPT are limited. Objective: This paper aimed to analyze ChatGPT’s performance on USMLE Step 3 practice test questions to better elucidate the strengths and weaknesses of AI use in medical education and deduce evidence-based strategies to counteract AI cheating. Methods: A total of 2069 USMLE Step 3 practice questions were extracted from the AMBOSS study platform. After including 229 image-based questions, a total of 1840 text-based questions were further categorized and entered into ChatGPT 3.5, while a subset of 229 questions were entered into ChatGPT 4. Responses were recorded, and the accuracy of ChatGPT answers as well as its performance in different test question categories and for different difficulty levels were compared between both versions. Results: Overall, ChatGPT 4 demonstrated a statistically significant superior performance compared to ChatGPT 3.5, achieving an accuracy of 84.7% (194/229) and 56.9% (1047/1840), respectively. A noteworthy correlation was observed between the length of test questions and the performance of ChatGPT 3.5 (ρ=–0.069; P=.003), which was absent in ChatGPT 4 (P=.87). Additionally, the difficulty of test questions, as categorized by AMBOSS hammer ratings, showed a statistically significant correlation with performance for both ChatGPT versions, with ρ=–0.289 for ChatGPT 3.5 and ρ=–0.344 for ChatGPT 4. ChatGPT 4 surpassed ChatGPT 3.5 in all levels of test question difficulty, except for the 2 highest difficulty tiers (4 and 5 hammers), where statistical significance was not reached. Conclusions: In this study, ChatGPT 4 demonstrated remarkable proficiency in taking the USMLE Step 3, with an accuracy rate of 84.7% (194/229), outshining ChatGPT 3.5 with an accuracy rate of 56.9% (1047/1840). Although ChatGPT 4 performed exceptionally, it encountered difficulties in questions requiring the application of theoretical concepts, particularly in cardiology and neurology. These insights are pivotal for the development of examination strategies that are resilient to AI and underline the promising role of AI in the realm of medical education and diagnostics.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Kung, T. H., Cheatham, M., Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., Maningo, J., et al. (2023) · PLOS Digital Health · doi:10.1371/journal.pdig.0000198 · Cited by 4,115 · cited 2×
“For instance, Hoch et al [11] reported that ChatGPT answered 57% of facial surgery board certification test questions correctly, while Kung et al [12] used a limited set of USMLE test questions (USMLE Step 1: 119; USMLE Step 2 CK: 102; USMLE Step 3: 122) and found that ChatGPT achieved performance…”
BackgroundKung, T. H., Cheatham, M., ChatGPT, Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., et al. (2022) · doi:10.1101/2022.12.19.22283643 · Cited by 4,115
Gilson, A., Safranek, C. W., Huang, T., Socrates, V., Chi, L., Taylor, R. A., & Chartash, D. (2023) · JMIR Medical Education · doi:10.2196/45312 · Cited by 2,078
ChatGPT passing USMLE shines a spotlight on the flaws of medical education
Mbakwe, A. B., Lourentzou, I., Celi, L. A., Mechanic, O. J., & Dagan, A. (2023) · PLOS Digital Health · doi:10.1371/journal.pdig.0000205 · Cited by 273
Hoch, C. C., Wollenberg, B., Lüers, J.-C., Knoedler, S., Knoedler, L., Frank, K., Cotofana, S., & Alfertshofer, M. (2023) · European Archives of Oto-Rhino-Laryngology · doi:10.1007/s00405-023-08051-4 · Cited by 154 · cited 2×
“For instance, Hoch et al [11] reported that ChatGPT answered 57% of facial surgery board certification test questions correctly, while Kung et al [12] used a limited set of USMLE test questions (USMLE Step 1: 119; USMLE Step 2 CK: 102; USMLE Step 3: 122) and found that ChatGPT achieved performance…”
BackgroundInfluentialSelf-citationThe opportunities and pitfalls of ChatGPT in clinical and translational medicine
Baumgartner, C. (2023) · Clinical and Translational Medicine · doi:10.1002/ctm2.1206 · Cited by 103
Cited by
Jin, H. K., Lee, H. E., & Kim, E. (2024) · BMC Medical Education · doi:10.1186/s12909-024-05944-8 · Cited by 85
“Our finding is similar to what a previous study showed regarding USMLE questions, where accuracy rates of 84.7% were observed for GPT-4 and 56.9% for its earlier version [56].”
ResultInfluentialLiu, M., Okuhara, T., Dai, Z., Huang, W., Gu, L., Okada, H., Furukawa, E., & Kiuchi, T. (2025) · International Journal of Medical Informatics · doi:10.1016/j.ijmedinf.2024.105673 · Cited by 73
ChatGPT’s Response Consistency: A Study on Repeated Queries of Medical Examination Questions
Funk, P. F., Hoch, C. C., Knoedler, S., Knoedler, L., Cotofana, S., Sofo, G., Bashiri Dezfouli, A., Wollenberg, B., Guntinas-Lichius, O., & Alfertshofer, M. (2024) · European Journal of Investigation in Health, Psychology and Education · doi:10.3390/ejihpe14030043 · Cited by 69
Self-citationWang, L., Li, J., Zhuang, B., Huang, S., Fang, M., Wang, C., Li, W., Zhang, M., & Gong, S. (2025) · Journal of Medical Internet Research · doi:10.2196/64486 · Cited by 51
Lin, Y., Luo, Z., Ye, Z., Zhong, N., Zhao, L., Zhang, L., Li, X., Chen, Z., & Chen, Y. (2025) · JMIR Medical Education · doi:10.2196/71125 · Cited by 58
“…is particularly evident in medical education, where students often rely on the answers provided by GAI when solving complex problems rather than relying on their logical reasoning and knowledge accumulation for analysis and resolution [35,39,40,50,55,69,70,72,74,75,77,98,99,124, 136,146-152].”
Jaleel, A., Aziz, U., Farid, G., Zahid Bashir, M., Mirza, T. R., Khizar Abbas, S. M., Aslam, S., & Sikander, R. M. H. (2025) · JMIR Medical Education · doi:10.2196/68070 · Cited by 25
Metadata
- Authors
- Knoedler, L., Alfertshofer, M., Knoedler, S., Hoch, C. C., Funk, P. F., Cotofana, S., Maheta, B., Frank, K., Brébant, V., Prantl, L., & Lamby, P.
- Venue
- JMIR Medical Education
- Publisher
- JMIR Publications Inc.
- Published
- 2024-01-05
- Volume
- 10
- Pages
- e51148
- ISSN
- 2369-3762 (ISSN-L)
- OpenAlex
- W4390617165
- Cited by
- 49
- References
- 6
- FWCI
- 3.85
- Type
- journal article
- Language
- en
- Open access
- gold
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.