In February 2026, Google published results showing that its Gemini 2.5 Ultra model had passed every major medical licensing examination in the United States, the United Kingdom, and the European Union — in many cases with scores that placed it in the top decile of human candidates. The results were peer-reviewed and independently verified. The medical community’s reaction was divided.
The examinations tested clinical knowledge, diagnostic reasoning, pharmacology, and medical ethics. They are designed to be difficult: pass rates for first-time human candidates typically range from 70 to 85 percent. Gemini 2.5 Ultra passed all of them on the first attempt.
“This is a calibration moment for medicine,” said Dr. Eric Topol, director of the Scripps Research Translational Institute and one of the field’s leading voices on AI. “We have to separate what the model knows from what it can do. It knows an extraordinary amount. Whether it can apply that knowledge safely in complex real-world cases is a different question.”
That distinction matters enormously. Medical licensing examinations test knowledge and reasoning but cannot test the tactile, observational, and relational dimensions of clinical medicine that experienced physicians rely on. The model has no way to hear a patient’s breathing, observe their gait, or pick up on the hesitations in their description of symptoms that sometimes indicate something is being left unsaid.
Where AI medical tools are having unambiguous impact is in radiology and pathology — disciplines where the task is pattern recognition in images, exactly the kind of problem machine learning excels at. Several health systems in the US and UK have deployed AI tools for screening mammograms and chest X-rays, with early results showing significant reductions in missed diagnoses.
Leave a Reply