Microsoft study says its AI diagnosed 80% of simulated ED cases, versus 47% for doctors

In a study published October 22, 2026 in Science, Microsoft researchers reported that their MAI Diagnostic Orchestrator (MAI-DxO), running on OpenAI's GPT-5, correctly diagnosed 80% of simulated, text-only cases built from past emergency-department admission records, versus 47% for generalist physicians.

What Microsoft's study found. According to the paper, MAI-DxO scored 80.3% on emergency-department cases from the MIMIC-IV-ED dataset, against 46.8% for 16 generalist physicians, about 1.7 times as accurate. On harder New England Journal of Medicine (NEJM) case records, Microsoft reports 83% for the system versus 36% for physicians with references and 22% without. The physicians were scored on held-out test sets: 56 NEJM cases from 2024-2025 and 71 MIMIC-IV-ED cases. All 15 authors are Microsoft employees, according to Unite.AI.

This was a simulation, not an emergency department. No real patients were involved and no physician was tested in a hospital. The cases were de-identified historical records turned into text-only encounters, where an AI "gatekeeper" model revealed findings only when asked. The physicians were generalists, mainly primary care and hospital doctors, not specialists or teams. Medical Economics reports the group included some internists and emergency physicians.

The doctors could search, but not use generative AI. In the open-book arm, physicians could use search and reference tools such as UpToDate and NICE guidelines, but not generative AI such as ChatGPT or OpenEvidence. In the closed-book arm, physicians could not use search engines or any outside resources.

Most of the gain comes from the underlying model. GPT-5 on its own scored 76.6% on the emergency-department cases, according to the preprint, close to the orchestrator's 80.3%. Microsoft's orchestrator adds a structured, step-by-step process on top of the model: asking questions, ordering tests and weighing cost before committing to a diagnosis. Microsoft says the gains 'persisted – and improved' with newer models including GPT-5.5, but TechCrunch reports that on ED cases GPT-5.5 scored about 85% with or without the orchestrator, while on NEJM cases the orchestrator lifted it from 72% to nearly 87%.

Cost savings depend on the configuration. Costs were simulated from standardized US price lists, with tests mapped to billing codes and each physician question priced at $300; they are not real spending. On NEJM cases, the headline 83% configuration averaged $5,430 per case, more than the $4,725 for physicians with references. Only budget-capped versions, such as 79.6% at $1,302, were both cheaper and more accurate than physicians. On ED cases, the unbudgeted system averaged $4,979 per case versus $3,400 for physicians, according to Medical Economics' reading of the Science paper.

In plain language. Microsoft built software that coordinates an AI model to work through a medical case the way a doctor might, one question or test at a time. On written versions of old hospital cases, it named the right diagnosis more often than individual general doctors with search and reference tools. That is a lab result, not proof it helps real patients. If you are a patient or clinician, nothing changes today; the system is not in clinical use.

Key terms.

• MAI-DxO: Microsoft AI's Diagnostic Orchestrator, software that directs a language model through a case step by step.
• MIMIC-IV-ED: a public, de-identified database of past emergency-department visits used for research.
• NEJM case records: difficult teaching cases published by the New England Journal of Medicine.
• Gatekeeper: an AI model that holds the case file and answers only what is asked.
• Prospective study: research that follows new patients forward in time, rather than reviewing past records.

How to read the claims.

1. Treat the 1.7x figure as applying only to simulated, text-only emergency-department cases.
2. Compare the system against GPT-5 alone (76.6%), not just against physicians.
3. Check which configuration a cost claim refers to; the most accurate setting was not the cheapest.
4. Read the paper's limitations: no imaging, no physical exam, only cases with a confirmed disease, and no assessment of treatment or safety.

How the claim spread. Mustafa Suleyman, CEO of Microsoft AI, posted from his personal account that the system was "1.7x more accurate than physicians in the ED, even when they have access to the web," and did fewer tests. The emergency framing overstates a simulated, text-only study, and we found no published test-count figure for the emergency cases. Satya Nadella shared the post, citing reduced testing costs.

What's next. Microsoft says it plans a prospective study with Beth Israel Deaconess Medical Center in which patients will interact with the system directly. The preprint itself says the results "do not yet firmly establish the clinical efficacy of MAI-DxO." Exact figures in the published Science version may differ slightly from the preprint. Verified against Microsoft's post, the preprint, Medical Economics and TechCrunch on October 22, 2026.