376 cases of rare genetic diseases. All unsolved — despite years of intensive diagnostics by specialists. Then OpenAI pointed its o3 Deep Research model at the data. Result: 18 new diagnoses where human experts had hit a wall. The study was just published in NEJM AI, conducted by the Manton Center at Boston Children’s Hospital together with Harvard and OpenAI.
What actually happened
The researchers fed the model anonymized clinical summaries of all 376 cases. Not raw sequencing data, but what geneticists typically work with: symptoms, prior findings, suspected diagnoses. The model then independently searched the literature, matched gene variants against known disease patterns, and produced evidence-linked hypotheses.
Here’s the crucial part: the model didn’t diagnose anyone. It generated suggestions that specialists then reviewed and confirmed. That distinction matters — and it’s exactly the right approach.
Results by cohort
A 4.8 percent hit rate might sound modest. But consider what we’re dealing with: cases where teams of human geneticists, neurologists, and other specialists had failed. Some for years.
The breakdown:
- Neurodevelopmental: 10 out of 100 cases solved
- Neuromuscular: 4 out of 61
- Sudden unexpected death in pediatrics: 2 out of 200
- Early psychosis: 2 out of 15
The standout story is Kyra. For nearly 20 years, her doctors searched for a diagnosis for her muscle weakness. The model identified myofibrillar myopathy — a diagnosis that finally brought clarity after almost two decades.
What comes next
OpenAI is taking this further. GPT-Rosalind, their specialized life-sciences model, is meant to enable even deeper analyses going forward. The OpenAI Foundation also awarded a research grant to the Manton Center to continue the collaboration.
My take
What impresses me about this study isn’t the technology. It’s the methodology. Nobody here claimed AI can replace doctors. Instead, a model was deployed specifically as a hypothesis generator — a kind of second opinion that systematically combs through literature no human could cover at that breadth.
18 families got an answer after years of waiting. That’s not marketing. That’s real. At the same time, I stay clear-eyed: 4.8 percent also means 95 percent of cases remain unsolved. AI is a tool here, not a miracle cure. But a remarkably useful one.
Sources: OpenAI: Diagnose rare childhood diseases, NEJM AI (June 2026)