Hidden flaws behind expert-level accuracy of multimodal GPT-4 vision in medicine
Open accessThe DOI matches its Crossref record; 23 of the 23 references it cites were found in a registry or an index.
Abstract
Recent studies indicate that Generative Pre-trained Transformer 4 with Vision (GPT-4V) outperforms human physicians in medical challenge tasks. However, these evaluations primarily focused on the accuracy of multi-choice questions alone. Our study extends the current scope by conducting a comprehensive analysis of GPT-4V’s rationales of image comprehension, recall of medical knowledge, and step-by-step multimodal reasoning when solving New England Journal of Medicine (NEJM) Image Challenges—an imaging quiz designed to test the knowledge and diagnostic capabilities of medical professionals. Evaluation results confirmed that GPT-4V performs comparatively to human physicians regarding multi-choice accuracy (81.6% vs. 77.8%). GPT-4V also performs well in cases where physicians incorrectly answer, with over 78% accuracy. However, we discovered that GPT-4V frequently presents flawed rationales in cases where it makes the correct final choices (35.5%), most prominent in image comprehension (27.2%). Regardless of GPT-4V’s high accuracy in multi-choice questions, our findings emphasize the necessity for further in-depth evaluations of its rationales before integrating such multimodal AI models into clinical workflows.
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
Figures


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.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“Recent studies on GPT-4V’s performance on medical multiple-choice questions demonstrated that despite the impressive results of models on multiple-choice questions, the rationale behind correct answers is flawed in a significant percentage of cases39.”
From Large Language Models lack essential metacognition for reliable medical reasoning (2025) · cites this paper as [39]
The cited source says what the claim attributes to it.
In the citing paper
“However, the current capabilities of LLMs, including advanced models like GPT-4V, remain limited in their understanding and processing of medical images56,57”
From A large language model digital patient system enhances ophthalmology history taking skills (2025) · cites this paper as [56]
The cited source says what the claim attributes to it.
In the citing paper
“AI-powered diagnoses33 are not yet perfected,34 and there is an ongoing discussion as to their use.35–39”
From Clinical AI Scribes in primary care: accuracy, error severity and implications for clinical practice (2025) · cites this paper as [34]
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
Large language models encode clinical knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2023) · Nature · doi:10.1038/s41586-023-06291-2 · Cited by 4,083 · cited 4×
“These re fl ect the multimodal nature of the task as well as the commonly evaluated capabilities for text-only LLMs 6,8 .”
BackgroundInfluentialLarge Language Models Encode Clinical Knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2022) · arXiv · doi:10.48550/arxiv.2212.13138 · Cited by 4,083 · cited 4×
“These re fl ect the multimodal nature of the task as well as the commonly evaluated capabilities for text-only LLMs 6,8 .”
BackgroundInfluentialAcosta, J. N., Falcone, G. J., Rajpurkar, P., & Topol, E. J. (2022) · Nature Medicine · doi:10.1038/s41591-022-01981-2 · Cited by 1,365
Capabilities of GPT-4 on Medical Challenge Problems
Nori, H., King, N., McKinney, S. M., Carignan, D., & Horvitz, E. (2023) · arXiv · doi:10.48550/arxiv.2303.13375 · Cited by 509 · cited 1×
“…fi ed by generative pre-trained transformer 4 (GPT-4) 1 have achieved remarkable performance on various biomedical tasks 2 , including summarizing medical evidence 3 , assisting in literature search 4,5 , answering medical examination questions 6 – 9 , andmatchingpatientstoclinicaltrials 10 .”
BackgroundEvaluating large language models on medical evidence summarization
Tang, L., Sun, Z., Idnay, B., Nestor, J. G., Soroush, A., Elias, P. A., Xu, Z., Ding, Y., Durrett, G., Rousseau, J. F., et al. (2023) · npj Digital Medicine · doi:10.1038/s41746-023-00896-7 · Cited by 423
Self-citationEvaluating Large Language Models on Medical Evidence Summarization
Tang, L., Sun, Z., Idnay, B., Nestor, J. G., Soroush, A., Elias, P. A., Xu, Z., Ding, Y., Durrett, G., Rousseau, J., et al. (2023) · doi:10.1101/2023.04.22.23288967 · Cited by 423
Self-citationOpportunities and Challenges for ChatGPT and Large Language Models in Biomedicine and Health
Tian, S., Jin, Q., Yeganova, L., Lai, P.-T., Zhu, Q., Chen, X., Yang, Y., Chen, Q., Kim, W., Comeau, D. C., et al. (2023) · Briefings in Bioinformatics · doi:10.1093/bib/bbad493 · Cited by 359
Self-citationCan large language models reason about medical questions?
Liévin, V., Hother, C. E., Motzfeldt, A. G., & Winther, O. (2024) · Patterns · doi:10.1016/j.patter.2024.100943 · Cited by 343 · cited 1×
“These re fl ect the multimodal nature of the task as well as the commonly evaluated capabilities for text-only LLMs 6,8 .”
BackgroundCan large language models reason about medical questions?
Liévin, V., Hother, C. E., Motzfeldt, A. G., & Winther, O. (2022) · arXiv · doi:10.48550/arxiv.2207.08143 · Cited by 343 · cited 1×
“These re fl ect the multimodal nature of the task as well as the commonly evaluated capabilities for text-only LLMs 6,8 .”
BackgroundMatching Patients to Clinical Trials with Large Language Models
Jin, Q., Wang, Z., Floudas, C. S., Chen, F., Gong, C., Bracken-Clarke, D., Xue, E., Yang, Y., Sun, J., & Lu, Z. (2023) · arXiv · doi:10.48550/arxiv.2307.15051 · Cited by 287 · cited 1×
“…fi ed by generative pre-trained transformer 4 (GPT-4) 1 have achieved remarkable performance on various biomedical tasks 2 , including summarizing medical evidence 3 , assisting in literature search 4,5 , answering medical examination questions 6 – 9 , andmatchingpatientstoclinicaltrials 10 .”
BackgroundSelf-citationMatching patients to clinical trials with large language models
Jin, Q., Wang, Z., Floudas, C. S., Chen, F., Gong, C., Bracken-Clarke, D., Xue, E., Yang, Y., Sun, J., & Lu, Z. (2024) · Nature Communications · doi:10.1038/s41467-024-53081-z · Cited by 287 · cited 1×
“…fi ed by generative pre-trained transformer 4 (GPT-4) 1 have achieved remarkable performance on various biomedical tasks 2 , including summarizing medical evidence 3 , assisting in literature search 4,5 , answering medical examination questions 6 – 9 , andmatchingpatientstoclinicaltrials 10 .”
BackgroundSelf-citationCan Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine
Nori, H., Lee, Y. T., Zhang, S., Carignan, D., Edgar, R., Fusi, N., King, N., Larson, J., Li, Y., Liu, W., et al. (2023) · arXiv · doi:10.48550/arxiv.2311.16452 · Cited by 178 · cited 1×
“…fi ed by generative pre-trained transformer 4 (GPT-4) 1 have achieved remarkable performance on various biomedical tasks 2 , including summarizing medical evidence 3 , assisting in literature search 4,5 , answering medical examination questions 6 – 9 , andmatchingpatientstoclinicaltrials 10 .”
BackgroundSelf-citationAs artificial intelligence goes multimodal, medical applications multiply
Topol, E. J. (2023) · Science · doi:10.1126/science.adk6139 · Cited by 132 · cited 1×
“However,mostoftheseLLMsare unimodal, utilizing only the free-text context, while clinical tasks often require the integration of narrative descriptions and multiple types of imaging tests 11,12 .”
BackgroundPubMed and beyond: biomedical literature search in the age of artificial intelligence
Jin, Q., Leaman, R., & Lu, Z. (2024) · eBioMedicine · doi:10.1016/j.ebiom.2024.104988 · Cited by 124
Self-citationZhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., et al. (2023) · arXiv · doi:10.48550/arxiv.2303.00915 · Cited by 101
Multimodal ChatGPT for Medical Applications: an Experimental Study of GPT-4VPreprint
Yan, Z., Zhang, K., Zhou, R., He, L., Li, X., & Sun, L. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2310.19061 · Cited by 95
Jin, Q., Leaman, R., & Lu, Z. (2023) · Journal of the American Society of Nephrology · doi:10.1681/asn.0000000000000166 · Cited by 64
Self-citationWu, C., Lei, J., Zheng, Q., Zhao, W., Lin, W., Zhang, X., Zhou, X., Zhao, Z., Zhang, Y., Wang, Y., et al. (2023) · arXiv · doi:10.48550/arxiv.2310.09909 · Cited by 43 · cited 1×
“Subsequent pilot studies have been conducted to analyze the performance of GPT-4V in the medical domain 13 – 16 (summarized in Supplementary Table 2).”
BackgroundYang, Z., Yao, Z., Tasmin, M., Vashisht, P., Jang, W. S., Ouyang, F., Wang, B., Berlowitz, D., & Yu, H. (2023) · medRxiv · doi:10.1101/2023.10.26.23297629 · Cited by 36 · cited 1×
“Our fi ndings, therefore, align with the previous ones, which show the expert-level performance of GPT-4V in the closed-book setting 15,16 .”
ResultMultimodal Foundation Models Exploit Text to Make Medical Image PredictionsPreprint
Buckley, T., Diao, J. A., Rajpurkar, P., et al. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2311.05591 · Cited by 18 · cited 2×
“Subsequent pilot studies have been conducted to analyze the performance of GPT-4V in the medical domain 13 – 16 (summarized in Supplementary Table 2).”
BackgroundResultMultimodal foundation models exploit text to make medical image predictions
Buckley, T. A., Diao, J. A., Srivastava, C. N., Brodeur, P. G., Rajpurkar, P., Rodman, A. M., & Manrai, A. K. (2026) · Nature Communications · doi:10.1038/s41467-026-74207-5 · Cited by 17
Encapsulating Peritoneal Sclerosis
Tang, X., & Sun, L. (2023) · New England Journal of Medicine · doi:10.1056/nejmicm2206513 · Cited by 4
OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., et al. (2023) · arXiv · doi:10.48550/arxiv.2303.08774
Self-citation
Cited by
Performance of a large language model on the reasoning tasks of a physician
Brodeur, P. G., Buckley, T. A., Kanjee, Z., Goh, E., Ling, E., Jain, P., Cabral, S., Abdulnour, R.-E. E., Haimovich, A. D., Freed, J. A., et al. (2026) · Science · doi:10.1126/science.adz4433 · Cited by 231
Large Language Models lack essential metacognition for reliable medical reasoning
Griot, M., Hemptinne, C., Vanderdonckt, J., & Yuksel, D. (2025) · Nature Communications · doi:10.1038/s41467-024-55628-6 · Cited by 172
“Recent studies on GPT-4V ’ s performance on medical multiple-choice questions demonstrated that despite the impressive results of models on multiple-choice questions, the rationale behind correctanswersis fl awedinasigni fi cantpercentageofcases 39 .”
Towards Injecting Medical Visual Knowledge into Multimodal LLMs at Scale
Chen, J., Gui, C., Ouyang, R., Gao, A., Chen, S., Chen, G. H., Wang, X., Cai, Z., Ji, K., Wan, X., et al. (2024) · Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2024.emnlp-main.418 · Cited by 74
Small language models learn enhanced reasoning skills from medical textbooks
Kim, H., Hwang, H., Lee, J., Park, S., Kim, D., Lee, T., Yoon, C., Sohn, J., Park, J., Reykhart, O., et al. (2025) · npj Digital Medicine · doi:10.1038/s41746-025-01653-8 · Cited by 59
The evaluation illusion of large language models in medicine
Agrawal, M., Chen, I. Y., Gulamali, F., & Joshi, S. D. (2025) · npj Digital Medicine · doi:10.1038/s41746-025-01963-x · Cited by 115
Metadata
- Authors
- Jin, Q., Chen, F., Zhou, Y., Xu, Z., Cheung, J. M., Chen, R., Summers, R. M., Rousseau, J. F., Ni, P., Landsman, M. J., Baxter, S. L., Al’Aref, S. J., Li, Y., Chen, A., Brejt, J. A., Chiang, M. F., Peng, Y., & Lu, Z.
- Affiliations
- National Institutes of Health (US)
- University of Pittsburgh (US)
- Weill Cornell Medicine (US)
- New York University (US)
- Harvard University (US)
- The University of Texas Southwestern Medical Center (US)
- MetroHealth Medical Center (US)
- University of California San Diego (US)
- University of Pittsburgh Medical Center (US)
- Corresponding author
- Peng, Yifan
- Lu, Zhiyong
- ORCID
- Venue
- npj Digital Medicine
- Published
- 2024-07-23
- Volume
- 7
- Issue
- 1
- Pages
- 190-190
- ISSN
- 2398-6352 (ISSN-L)
- PMC
- PMC11266508
- OpenAlex
- W4400921803
- Cited by
- 146
- References
- 18
- Citations by year
- 2026: 67 · 2025: 60 · 2024: 19
- FWCI
- 13.65
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- npj Digital Medicine (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- eScholarship (California Digital Library) (submittedVersion, cc-by)
- PubMed Central (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Funders
- U.S. Department of Health & Human Services | National Institutes of Health
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Keywords
- GPT-4V, multimodal large language models, medical image interpretation, diagnostic accuracy, multimodal reasoning, clinical implementation of AI
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.