A new adversarial stress test in Nature Medicine shows leading multimodal models still lean on shortcuts, fabricate details, and falter under small input tweaks even when they ace ordinary health benchmarks, leaving a clear gap between lab scores and actual clinical reliability.
We stress tested many frontier AI models for multimodal medical reasoning (including GPT-5, Claude 3.5, Gemini 2.5 Pro). They’re not ready. Faulty reasoning, use of inappropriate shortcuts, hallucinations. Published today @NatureMedicine https://www.nature.com/articles/s41591-026-04501-8
Simple changes like removing key image details or swapping modalities exposed how models reach correct answers for the wrong reasons, a pattern the study ties directly to current test design rather than model size alone.
The authors released perturbation tools and rubrics for others to reuse, yet no model-specific scores or vendor replies appear in the published record, so any timeline for fixes remains unknown for now.
Many users dismissed the Nature Medicine study on frontier AI models failing medical reasoning tests as meaningless or worthless because it used outdated models, while some praised its rigor or expressed hope for future progress.
No Digg Deeper questions have been answered for this story yet.
Link for free access https://rdcu.be/fqznS This extensive assessment work was led by @hoifungpoon and Yu (Aiden) Gu
We stress tested many frontier AI models for multimodal medical reasoning (including GPT-5, Claude 3.5, Gemini 2.5 Pro). They’re not ready. Faulty reasoning, use of inappropriate shortcuts, hallucinations. Published today @NatureMedicine https://www.nature.com/articles/s41591-026-04501-8
tests 5-generation-old models, concludes AI is inappropriate for medicine
We stress tested many frontier AI models for multimodal medical reasoning (including GPT-5, Claude 3.5, Gemini 2.5 Pro). They’re not ready. Faulty reasoning, use of inappropriate shortcuts, hallucinations. Published today @NatureMedicine https://www.nature.com/articles/s41591-026-04501-8
Many people have already pointed out that no matter how high the quality of this paper, the long review and publication cycle makes the results irrelevant.
The way to fix this is to open source enough of the actual research methodology used so that upon publication, anyone can re-run the exact same tests on the latest models at that time to produce consistently comparable results.
This is more useful because “this is the worst the models will ever be,” and it is a reasonable assumption that they at some point WILL be ready. Hence, showing that at some point in time that they aren’t (i.e. a year ago) is much less useful than constructing a usable method that allows us to tell at what point in the future they actually ARE ready.
We stress tested many frontier AI models for multimodal medical reasoning (including GPT-5, Claude 3.5, Gemini 2.5 Pro). They’re not ready. Faulty reasoning, use of inappropriate shortcuts, hallucinations. Published today @NatureMedicine https://www.nature.com/articles/s41591-026-04501-8
@EricTopol @NatureMedicine It does take time, and I'm not suggesting that your study lacked rigor. The problem is that the target is moving too quickly to be evaluated this way. The results, once finally published, are misleading because they don't represent the state of the art.
@Timothy98537991 @NatureMedicine Obviously from someone who has never tried to publish such as assessment in a leading peer review journal. It takes time! Best we can do.
@EricTopol @NatureMedicine I learned way too late in life that the truth can defend itself and doesn't need me to defend it. I think it will be pretty obvious to most people that this study was accurate at the time it was done, but by the time it was published, it was meaningless.
@EricTopol @NatureMedicine Those are not frontier models
@SandelloRed @NatureMedicine Yes they are
@EricTopol @Timothy98537991 @NatureMedicine Yes, but you could have tweeted "they were not ready" rather than using the present tense.
@EricTopol @NatureMedicine Would it be possible to rapidly test the same stuff again with the up-to-date models?
I'm not expecting perfection from GPT-5.5 or Opus 4.8 but maybe relevant improvement, still.
@EricTopol @NatureMedicine @josegallucci
@EricTopol @NatureMedicine @alinefortuna2
@EricTopol @NatureMedicine Maybe be prepared with personal results of the same tests against current models to supplement your paper’s findings. Otherwise it is fairly meaningless to post, especially with your present tense framing of “they’re not ready.”
@EricTopol @NatureMedicine The study is outdated now, get ready for GPT6, claude mythos, Gemini 3.5 PRO
Or better, wait for a year
@EricTopol @NatureMedicine I love this kind of research Eric, thank you for doing this. How do you suggest we mitigate the publishing delay issue for updating the research on newer models?
@ESRogs @Timothy98537991 @NatureMedicine The ones we tested were not ready. I indicated some in the text of the post. You’re welcome
@EricTopol @NatureMedicine The journal process is too slow for evaluating LLMs. Journals are not appropriate tools for this task.
@Timothy98537991 @EricTopol @NatureMedicine Hi Dr. Topol! Huge fan of your work, but I have to agree with Timothy on this one. We need to evolve as a field. While important, the traditional peer review process is simply far too slow to accurately assess the rapid pace of these models. (1/3)
@DrAMansouri @EricTopol @NatureMedicine You're right, there are two benchmarks. Absolute perfection vs "as good as humans". Other studies suggest models are already as good or better than humans in many medical contexts.
@yishan @NatureMedicine Agree. They will eventually be ready!
Many people have already pointed out that no matter how high the quality of this paper, the long review and publication cycle makes the results irrelevant.
The way to fix this is to open source enough of the actual research methodology used so that upon publication, anyone can re-run the exact same tests on the latest models at that time to produce consistently comparable results.
This is more useful because “this is the worst the models will ever be,” and it is a reasonable assumption that they at some point WILL be ready. Hence, showing that at some point in time that they aren’t (i.e. a year ago) is much less useful than constructing a usable method that allows us to tell at what point in the future they actually ARE ready.
