Microsoft MAI-DxO Science Study Tops Doctors, Faces FDA Gap

What happened

On October 22, 2026, Microsoft's diagnostic AI went through peer review. The company published its MAI Diagnostic Orchestrator (MAI-DxO) results in Science, building on work it first released as a preprint in June 2025.9 Running on OpenAI's GPT-5, the system reached the correct diagnosis in 83.2% of 304 complex case records from The New England Journal of Medicine. Generalist physicians scored 36.1% with search and reference tools and 22.3% without them.41 On a second set of 325 emergency department admissions, MAI-DxO scored 80.3% and physicians scored 46.8%.42

Microsoft AI CEO Mustafa Suleyman called the result a step toward "medical superintelligence."45 The company also announced a partnership with Beth Israel Deaconess Medical Center in Boston for a prospective study. Patients will interact with the system directly, and its conclusions will be checked against physicians' independent assessments.9 Microsoft has not given a start date for that study or any timeline for a release.49 The system is still a research tool with no product, API or clinical approval. In October, Microsoft released its benchmark, SDBench, as open source.44

How the system works

Most diagnostic AI reads a complete case file and returns an answer in one pass. MAI-DxO starts with a short vignette and has to ask for every further detail itself.42 Separate AI roles maintain a ranked list of likely diagnoses, choose tests that can tell them apart, and push back against anchoring bias. They then decide together whether to ask a question, order a test or commit to a diagnosis.4 A "gatekeeper" model releases case information only when asked. If a requested finding is not in the record, it makes up a normal result so the system can't infer the answer from gaps in the file.342

The design does not depend on one company's model. Microsoft tested it on systems from OpenAI, Google, Anthropic, Meta and DeepSeek, and OpenAI's GPT-5 and GPT-5.5 produced the best scores.45 The system also tracks spending: users can set a test budget, and every question and test is priced.42

Where the coverage agrees, and where it splits

Outlets broadly agree on the headline numbers and the main limits. The paper itself says its comparisons "do not establish superiority over specialists, multidisciplinary teams, or real-world clinical practice." The cases were text-only with no imaging, none involved a healthy patient, and the study measured diagnosis rather than treatment or outcomes.41 Several reports point out that cases this hard normally go to specialists or teams of doctors, and neither group was tested.4345 One critic notes that every author is a Microsoft employee with equity in the company, no outside funding was involved, and eight authors are named on a Microsoft patent application covering the evaluation framework.50

The gap between doctors and the AI has also shrunk since the original announcement. In June 2025, Microsoft said the best setup reached 85.5% while 21 experienced physicians averaged 20%, roughly four times worse.1 Even then, observers pointed out that physicians usually rely on search engines and generative AI at work, so blocking those tools probably pushed their scores down.7 The Science version tests doctors both with and without references. The physicians in the preprint had been barred from online sources.41 With references allowed, the AI's lead falls from about four-to-one to a little more than two-to-one. That is still a large gap, but it is a more defensible claim than the original one.

The biggest disagreement in the coverage is about how much the orchestrator adds and what it costs. One analysis says GPT-5 alone scored 76.6% at about $4,240 per case. Unconstrained MAI-DxO on GPT-5 scored 80.3% at about $4,980, a gain of 3.7 points that cost roughly $740 more per case.9 Another report lists GPT-5 alone at 78.9% against MAI-DxO's 83.2%.43 Another account says the orchestrator gave GPT-5 about four points and roughly $3,000 in savings per case. On the hard cases it lifted GPT-5.5 from 72% to nearly 87% and Gemini 2.5 Flash from 46.4% to nearly 69%, but neither GPT-5 model improved meaningfully on emergency cases.46 On costs, a budget-capped setup reportedly reached 79.6% at about $1,303 per case, compared with $4,725 for physicians using references.42

These numbers likely come from different configurations and test sets, so the conflict may be smaller than it looks. Read together, they suggest the base model is doing most of the work. The orchestrator helps weaker models most and gives top models a smaller lift. Science coverage reached the same conclusion: stronger base models drive much of the improvement, and orchestration adds gains on top.45 For hospitals, that matters. An advantage built on someone else's model could disappear when the next model arrives.

The system's behavior also stands out. It asked an average of 25.5 questions before ordering its first test, compared with 5.6 for physicians.9 That approach works well on a benchmark. In a busy emergency department, it would mean a lot of history-taking.

The regulatory picture

Regulation is where the "superintelligence" framing runs into reality. As of September 2026, the FDA says it has authorized more than 1,600 AI-enabled medical devices.31 An analysis of the September 4 list counts 1,614, with 96.2% cleared through the 510(k) pathway, which requires showing a device is substantially equivalent to one already on the market. Radiology accounts for 76.2% of them.39 Narrow tools dominate the list: image segmentation, triage software and ultrasound features.35 As of March 2026, no device using generative AI or large language models had been authorized.32 Coverage of the Science paper says no generative AI diagnostic system has FDA clearance, and that the agency's January 2025 draft guidance on AI device lifecycles has still not been finalized.42

The FDA's main policy move this year points in two directions. Its January 6, 2026 clinical decision support guidance, which Commissioner Marty Makary described as cutting unnecessary regulation, says the agency will generally not enforce device rules on software that gives a single recommendation, as long as clinicians can independently review the basis for it.14 Lawyers say this lowers the barrier for some AI copilots aimed at clinicians. They also caution that opaque, time-critical or directive models should still expect device oversight.13 Several firms note that despite how the announcement was framed, the guidance text says nothing about AI and does not address consumer-facing chatbots or symptom checkers.16 The FDA treats critical, time-sensitive tools such as stroke and sepsis alarms as generally failing the independent-review test.18 Software aimed at patients or caregivers is still regulated as a device.19 Separately, the FDA released a discussion paper on generative AI devices in August, with comments due October 19. It treats patient-facing generative AI as potentially higher risk and warns about conversations that drift from information toward directing action.19

Put MAI-DxO against those rules and the conclusion is fairly clear. The Beth Israel study has patients interacting with the system directly, beginning with emergency department triage.42 Patient-facing use and time pressure are the two features most likely to push a product out of the lighter regulatory lane and into full device review. Any commercial version would probably face the toughest scrutiny the FDA applies, not the relaxed approach of the January guidance.

What hospitals are actually deploying

Hospital use of AI has largely moved past diagnosis toward paperwork. Microsoft's DAX Copilot documentation tool runs in more than 150 health systems, while diagnostic AI built on language models remains mostly in research.24 One industry tally says about 80% of hospitals use AI somewhere, but fewer than 20% have it built into core clinical diagnosis.21 Deployed diagnostic tools tend to be narrow and validated, such as Johns Hopkins' sepsis algorithm. It reported 82% sensitivity and an 18% relative drop in mortality when clinicians acted on alerts within an hour.29 The risks are also real. An Ontario auditor found AI transcription tools putting made-up content into physician notes.29

The competition is getting crowded. Google tested its AMIE chatbot at the same Boston hospital.41 Suleyman says Microsoft wants to lead in clinical decision support products over the next few years.46

The verdict

The Science paper is a real step forward. It is a stronger benchmark that tests step-by-step reasoning and cost, and it gives physicians a fairer comparison than the preprint did. But it shows the system is good at puzzles. It does not show the system is a better doctor. The figures that will decide whether this becomes medicine are not in the paper yet: how the system does with real patients, how often clinicians ignore its advice, and how often it raises false alarms in people who turn out to be healthy. Meanwhile the FDA has made room for transparent tools that keep clinicians in charge, while leaving the patient-facing generative AI that Microsoft is about to test in a regulatory gray zone. The Beth Israel results will show whether the system works with real patients, and they will matter more than another benchmark score.