Microsoft Research podcast: Horvitz and Topol on AI in medicine

Microsoft Research has published an episode of its podcast series The Shape of Things to Come, hosted by Microsoft Research leader Doug Burger. The series aims to dig into the fundamentals of AI capabilities and think through how they will change the world, for better and worse. This episode asks what medicine might look like with AI, and what needs to happen first. Burger's guests are Eric Horvitz, Microsoft Chief Scientific Officer; Eric Topol of Scripps Research, a physician-scientist and Microsoft AI health advisor; and Ava Amini, Microsoft Principal Research Manager. The episode description says the group covers AI-assisted diagnosis, medical imaging, whole-patient models, virtual cells and biological foundation models, plus why research is slow to reach patients and why better biological data and real-world validation remain essential.
Burger introduces the guests. Horvitz is both a PhD and an MD. Topol is executive vice president of Scripps Research, which Burger calls the largest nonprofit biomedical research institute in the United States, and he has published over 1,300 peer-reviewed papers and authored four books. Burger notes that Topol is the one person on the podcast who is not a Microsoft employee. Amini is an expert in AI and precision medicine and co-lead of Project Ex Vivo, a collaboration between Microsoft and the Broad Institute on defining, engineering and targeting cell states in cancer. Burger opens by saying he has seen recent talks in which scientists with cancer used AI to design drugs for themselves.
Horvitz speaks first. He says healthcare and the biosciences are fortunate because their content has been digitalized, from protein structure over decades to electronic health records, which makes them amenable to AI. He is extremely optimistic about the long-term future. He points to progress over the last five to eight years with both discriminative and generative AI models, calls AlphaFold the tip of the iceberg, and sees a wave of possibilities in experimental platforms, experimental design, data interpretation and hypothesis generation, with AI as an assistant to human scientists. In healthcare he names patient safety, timely diagnosis and access as long-term challenges.
He then describes his team's work on diagnostic benchmarking, which he says moved from giving an AI a whole case and scoring its differential diagnosis to what he calls medical cognition. In the sequential diagnosis challenge the AI takes the clinician's place. It starts from a chief complaint and some history, then chooses which questions to ask and which tests to order, balancing the information value against the cost of tests, and decides when to call a diagnosis, even under uncertainty. Horvitz says this showed remarkable abilities in a number of frontier language models. He adds that the comparisons with human generalist physicians have limitations and must be taken at face value, but still show promise for assisting future clinicians. Burger congratulates him on the paper, which he describes as a cover article in Science.
Topol also congratulates Horvitz and the Microsoft team. He calls the Science paper seminal and says it was the first to use an agentic system to make more accurate diagnoses than physicians and than many other frontier models. In his words, accuracy improved markedly over physicians, almost doubling, at lower cost. No baseline or percentage is given. He contrasts this with AI used to upcode and raise costs, and says it was the first demonstration of higher accuracy at lower cost.
Topol's broader point is that the field has fixed on generative AI, which he says is four years in, and has missed where AI is even more proven: deep learning on medical images such as the retina, chest x-ray and electrocardiogram. He calls it a paradox that what has been learned from mammography, including a randomized trial of over a hundred thousand women, and from colonoscopy, with 44 randomized trials, is not being implemented. He says AI can read information embedded in every type of medical image that humans cannot see.
He names whole-patient models as the next phase of AI. They would combine electronic health records, raw images and other data layers such as genomics, proteomics and the exposome, and would let clinicians predict diseases and when they might occur and test what-if scenarios for deferring or preventing them. But, he says, the work done so far has not affected patient care, which he finds vexing.
Asked what the community should do differently, Topol answers that it is missing implementation. The evidence sits in peer-reviewed journals, including randomized trials run in the real world, and goes unused. He sees no good reasons, only excuses: the medical profession is slow to adopt change, there are financial constraints, and health systems working at very low margins are unwilling to invest in things they are not fully confident about. He adds that neither physicians nor patients are clamoring for these advances, and says almost no women in the United States get state-of-the-art AI in mammography. The visible transcript ends partway through his next answer.
Key facts
- Microsoft Research's podcast The Shape of Things to Come, hosted by Doug Burger, features Eric Horvitz, Eric Topol of Scripps Research and Ava Amini on AI and the future of medicine.
- Horvitz describes a sequential diagnosis challenge in which an AI plays the clinician, choosing questions and tests and deciding when to call a diagnosis; he says frontier language models showed remarkable abilities, with comparisons to generalist physicians that carry limitations.
- Topol calls the related Science paper the first to use an agentic system for diagnoses more accurate than physicians' and many frontier models', with accuracy 'almost doubling' at lower cost; he gives no baseline.
- Topol says proven image AI goes unused: he cites a randomized mammography trial of over a hundred thousand women and 44 colonoscopy randomized trials as lessons not being implemented.
- He names whole-patient models, combining health records, raw images, genomics, proteomics and the exposome, as the next phase, and says the gap is implementation, not research.
Why it matters
The episode puts two views side by side. Horvitz is optimistic about AI because healthcare and the biosciences are already digitalized, and he describes a diagnostic test in which AI has to ask questions, order tests and decide when to commit, not just score a finished case. Topol praises that work as the first agentic system to beat physicians on accuracy at lower cost, then argues the real bottleneck is that proven AI, especially in medical imaging, does not reach patients. The discussion is a conversation, not a new release or result.
Who it affects
Patients, clinicians and health systems are the focus. Topol says women getting mammography in the United States almost never receive state-of-the-art AI interpretation, and that health systems work at very low margins and are reluctant to invest. The conversation is also aimed at technologists, policymakers and business decision makers, according to the series description, and at a broad listenership that is not limited to AI or medicine experts.
How to use it
There is no product, tool or price in the episode. It is a listening resource in the Microsoft Research Podcast series, useful as an overview of how Microsoft researchers and an outside physician-scientist frame AI in diagnosis, imaging and whole-patient models. Readers who want the diagnostic results should look to the Science paper Horvitz and Topol discuss.
How solid is it
This is an interview, so the claims are the speakers' own. Topol's 'first ever' and 'almost doubling' statements about the Science paper come with no baseline, percentage or cost figure in the visible text, and Horvitz says the comparison with human generalist physicians has limitations and should be taken at face value. Topol is a Microsoft AI health advisor, and Burger and the guests congratulate each other on the paper. The visible text is truncated mid-answer in Topol's second response, so Amini's remarks and the discussion of virtual cells, drug discovery and biological data are not visible.
Risks and caveats
Horvitz's comparison with physicians is described by him as limited, and the models tested and the number of physicians are not named in the visible text. Topol's whole-patient models are a prospect, not a deployed system, and no timescale is given for them or for wider adoption of AI imaging. His own account of the obstacles is slow adoption, financial constraints and low appetite for investment, and he says physicians and patients are not pushing for these advances. The episode description itself says better biological data and real-world validation remain essential.
“This really was the first demonstration you could have higher accuracy at lower cost.”
— Eric Topol, Scripps Research, on the Science paper by Eric Horvitz and Microsoft colleagues