Medical AI reads the test, not the image
A Nature Medicine study by Eric Topol and Microsoft Research found GPT-5 scores 41.32% on image-dependent NEJM diagnostic tasks even with the image removed (random baseline: 20%), because the model exploits question wording and answer structure rather than reading pixels.
GPT-5 scored 41.32% on the subset of NEJM Image Challenge tasks where the image was necessary to answer, with the image removed. NEJM Image Challenge is a diagnostic quiz from the New England Journal of Medicine: physicians see a clinical image and choose from five diagnoses. Random guessing gives 20%. The 21-point gap means the model is pulling answers from question wording, answer choices, and memorized test patterns.
On 26 June, Nature Medicine published a study by Eric Topol of Scripps Research and a team at Microsoft Research. They tested multimodal models (GPT-5, Gemini 2.5 Pro, OpenAI o3, and others) on tasks where the system must read both patient text and a clinical image: X-ray, skin photo, retinal scan. On standard leaderboards, GPT-5 reached 81.33% on NEJM Image Challenge. The authors designed stress tests: removed the image, shuffled answer order, replaced wrong choices with plausible alternatives, swapped in an image pointing to a different diagnosis. The question: does the model read pixels or exploit the textual structure of the task?
Without the image, Gemini 2.5 Pro scored 40.10%, o3 scored 38.58%. All three scored roughly double the random baseline, which points to heavy reliance on question text.
GPT-4o scored 16.35% without the image, lowest in the group. The model often refused to answer when the image was absent. In a clinical setting, that refusal may be more useful than a confident guess: a physician needs a system that recognizes missing data.
Second stress test: the original image was replaced with one pointing to a wrong diagnosis. GPT-5 fell from 84% to 53%, o3 from 83.5% to 50.5%, Claude Sonnet 3.5 from 59% to 27.5%. Even with the swapped image, GPT-5 and o3 answered correctly more than half the time, leaning on textual context.
Chain-of-thought reasoning did not help. Across 120 NEJM tasks, asking models to "think step by step" slightly worsened results for GPT-5, Gemini 2.5 Pro, o3, and o4-mini. The authors separately documented false reasoning: a model picks a diagnosis without the image, then describes "observed" features that were never in the input, in confident clinical language. That confidence is the danger: the error reads like a professional conclusion.
Microsoft Research profile on Eternal Search.
The authors' conclusion: a model must ground its answer in the actual image, refuse when data are missing, and change its diagnosis when the visual evidence changes. Without that kind of stress test, a high accuracy number on imaging benchmarks says little about whether the model actually reads the image.
Sources
nature.com
t.me
The item provides concrete, quantified evidence that benchmark scores for top medical AI models overstate their image-reading ability, with Microsoft Research named as a co-author of the proof; Microsoft Research is tracked on Eternal Search as a biomedical research and funding organization, giving readers a direct way to follow their broader output beyond this single paper.