Two new medical AIs match or exceed doctors on diagnosis — while keeping clinicians in charge

Diagnosing a patient is one of the hardest things a doctor does: it means weaving together a history, symptoms, test results and clinical judgment under time pressure. In a study published in Nature on June 18, 2026, researchers presented two independent artificial-intelligence systems that can assist across several stages of that work — and, in careful evaluations, performed at least as well as physicians.

The first system, MIRA (Medical Intelligence for Reasoning and Action), was developed by Jakob Kather and colleagues. Across a set of clinical cases, MIRA reached an average diagnostic accuracy of 87.8 percent, compared with 78.1 percent for a panel of six physicians drawn from different specialties. The second, Google’s AMIE (Articulate Medical Intelligence Explorer), performed as well as real physicians in management reasoning and, the authors reported, better than physicians in the precision of the treatments and investigations it recommended.

“In a study published in Nature on June 18, 2026, researchers presented two independent artificial-intelligence systems that can assist across several stages of that work — and, in careful evaluations, performed at least as well as physicians.”

What makes the results encouraging rather than alarming is how the researchers frame the technology. These are decision-support tools built to work alongside clinicians who retain authority over every choice. A system that can quickly synthesize a complex case, flag a diagnosis a tired team might miss, or double-check a proposed treatment is most valuable as a second opinion — one that never gets fatigued and can draw on a vast base of medical literature in seconds.

The potential reach is significant. Much of the world faces a shortage of specialists, and patients in rural or under-resourced areas can wait weeks or months for expert review. A reliable assistant that helps a general practitioner reason through a difficult case could shorten that gap, especially where access to specialists is limited.

The scientists are also clear about the caveats. The evaluations were conducted on curated cases and simulated scenarios, not yet in the messy reality of everyday clinics, and moving from benchmark performance to safe bedside use will require prospective trials, regulatory review and close attention to how the tools behave across diverse populations. Published in one of the world’s most rigorous journals, the work is a measured, hopeful signal: not a robot doctor, but a tireless partner that could help human clinicians give more people careful, well-reasoned care.