Microsoft’s new MAI-DxO system beat human doctors at diagnosing complex medical conditions in a recent study. While doctors got the right answer 36% of the time, the AI reached 83% accuracy. Dominic King said the results were "very substantial and exciting" as the tool helps address the medical expert shortage.
Microsoft is taking a massive leap in healthcare. An AI system from the tech giant has proven to be more than twice as accurate as human doctors at diagnosing complex medical conditions. According to a new study published in the journal Science, Microsoft AI CEO Mustafa Suleyman believes this is a major step toward building what he calls "medical superintelligence."
How does the AI doctor work?
The new system is called MAI-DxO (short for Microsoft AI Diagnostic Orchestrator). Unlike a normal chatbot, it works like a virtual panel of doctors. The AI uses multiple "agents" that team up to ask patients questions, order lab tests, and debate symptoms before agreeing on a final diagnosis.
In a simulation using tough medical cases from the New England Journal of Medicine, MAI-DxO got the right answer 83% of the time. On the other hand, general physicians who used medical reference tools were correct only 36% of the time.
There is a small catch: cases this difficult are usually handled by a team of medical specialists, but they were not part of this comparison. Still, Dominic King, Vice President of Health at Microsoft AI, called the results "very substantial and exciting," noting that the biggest challenge in healthcare right now is the shortage of medical experts.
The power of GPT-5 and GPT-5.5
With these results, Microsoft joins companies like Google and OpenAI in claiming their AI can match or even beat human doctors.
The study showed that much of the AI's power comes from the advanced models running in the background. Microsoft tested models from OpenAI, Google, Meta, and others, but got the best results using OpenAI's GPT-5 and the newer GPT-5.5.
The DxO system also helps lower costs. For example, using GPT-5 alone on a complex case cost about $8,400 with a 78.9% accuracy. By adding the DxO system, Microsoft cut that cost by around $3,000 and slightly improved the accuracy.
With the newer GPT-5.5 model, DxO boosted accuracy from 72% to nearly 87% on complex cases, while keeping costs low. It also greatly helped less powerful AI models. Google's Gemini 2.5 Flash scored just 46.4% on its own, but jumped to nearly 69% when using Microsoft's DxO system. However, on fast-paced emergency department cases, the DxO system didn't show much improvement in accuracy or cost for the advanced models.
What happens next?
Suleyman, who previously co-founded Google’s DeepMind, has spent his career trying to solve medical issues with AI. He expects AI healthcare models to get incredibly better in the next two to three years. Microsoft is already working with the Mayo Clinic to build a dedicated AI model for healthcare, which is expected to be released next year. Suleyman told TechCrunch that by 2027, doctors will have highly accurate AI predictions about how a patient's medical record might unfold.
Before it can be used on real patients, Microsoft is putting DxO to the test in the real world. The company is partnering with the Beth Israel Deaconess Medical Center in Boston. During this trial, patients will receive standard care from human doctors. At the same time, the AI will silently run in the background. Researchers will then compare the AI’s recommendations with what the human doctors actually did.
"This is the kind of technology that everybody in the world wants," Suleyman said, promising that Microsoft will do everything it can to safely bring it to hospitals globally.