AI achieves high diagnosis accuracy

Microsoft's AI system, MAI-DxO, has been reported to correctly diagnose 83.2 percent of 304 complex medical cases, far outperforming general practitioners who achieved 36.1 percent accuracy using manuals and the Internet. The study was published in the journal Science on October 22, 2026, and was led by Mustafa Suleyman, head of Microsoft's AI division and co-founder of DeepMind.

AI functions as a diagnostic conductor

The AI system, described as a "diagnostic orchestrator," operates by controlling an existing AI model like GPT-5 from OpenAI and assigning it multiple roles. When presented with a fictional patient case, it begins with a brief summary and then queries a second program called the "Gatekeeper" to obtain additional information only when explicitly requested.

Cost and accuracy trade-off

According to simulated hospital costs in the United States, the most accurate version of MAI-DxO, which achieves 80.3 percent accuracy, costs $4,979 in tests per patient. A more economical version, costing $2,634, drops to 73.2 percent accuracy. General practitioners and internists, using manuals and the Internet, peak at 36.1 percent accuracy but drop to 22.3 percent without these tools.

AI asks more questions than doctors

The AI system asks an average of 25.5 questions before prescribing its first test, compared to 5.6 for doctors. These results were achieved in simulation, and no specialists or multidisciplinary teams participated in the study. The research excluded real-world complexities such as imaging, pathology slides, and patients who forget symptoms.

AI benefits modest models more

TechCrunch notes that GPT-5 alone, without the orchestrator, already reaches 78.9 percent accuracy. The orchestrator significantly improves the performance of less powerful models. For example, Gemini 2.5 Flash improves from 46.4 percent to 68.8 percent, and Claude 4.5 Sonnet improves from 69.4 percent to 81.9 percent. These models were all developed in 2025 and have since been updated.

AI continues to improve post-peer review

By the time the study was submitted to Science on April 4, 2026, the AI had advanced further. When tested with GPT-5.5, released in April, the orchestrator achieved nearly 87 percent accuracy on cases from the New England Journal of Medicine. In emergency rooms, GPT-5.5 alone matched the AI's performance, achieving around 85 percent accuracy.

Real-world testing planned

The technology is now ready for real-world testing. A trial is planned at Beth Israel Deaconess Medical Center in Boston, where volunteer patients will receive their usual care while also interacting with the AI. The AI's opinions will be compared to those of caregivers, though no specific date has been set for the trial.

Caution urged due to past errors

Caution is still needed, as a February 2026 study in Nature Medicine highlighted errors in ChatGPT Health during emergency situations. In Europe, such a tool would also require a CE marking for medical devices before being used with patients.