Illustration: TheHumanTouch.ai
The first signs were small. When Kyra was nine, she could no longer sink as deep into her karate stance, and she had slowed down on the soccer field. By 13 she was in a wheelchair and on a ventilator. For nearly two decades, nobody could tell her family why.
She is one of a great many patients whose genetic tests come back without an answer. There are thousands of rare diseases, the clues are buried in long lists of genetic variants and the medical literature changes every week. "The bottleneck is time," says Catherine Brownstein of the Manton Center for Orphan Disease Research at Boston Children's Hospital. "An expert can devote only so much of their day to any one particular person."
So the Manton Center, working with Harvard and OpenAI, tried something new. The team gave OpenAI's o3 deep-research model 376 cases that specialists had already studied and failed to crack, each stripped of names and packed with symptoms, doctors' notes and a filtered list of genetic variants. For each one, the model searched the literature and proposed the most likely explanation, showing its reasoning.
Then the humans took over. At least two experts reviewed every lead against the standard rules for judging genetic variants. Nothing counted as a diagnosis until it had been confirmed in a certified laboratory and the clinical team had told the family. The AI did not diagnose anyone. The doctors did.
The results, published in NEJM AI in June, were 18 new diagnoses, roughly one for every 20 cases that had been written off as unsolved. The yield was highest among children with developmental and neuromuscular disorders. In one case the model spotted a missing piece of chromosome 22 that did not appear anywhere in the data it was given; later sequencing proved it right. "Researchers like Catherine and me can't possibly keep 8,000 different diseases in our heads," says Alan Beggs, the centre's director. "That's the power of AI."
For Kyra, the answer was a small error in a gene called HSPB8, which causes a rare muscle disease called myofibrillar myopathy. She learned it about a week before her 28th birthday. A name does not undo her illness, which has since stopped getting worse. But after nearly 20 years of not knowing, she and her family finally know what they are facing.
The study has real limits. It looked back at old cases, so it cannot say how well the approach would work in everyday care, how much time it saves or how many false leads doctors would have to chase. The authors say larger studies are needed, and they stress that the work does not endorse using AI tools to diagnose yourself. Still, for families who are waiting, it carries a hopeful message: the answer may already be sitting in data that has been collected but that nobody has had time to read.
Kyra's story is drawn from OpenAI's account of the study, which does not give her surname. Sources: NEJM AI (June 2026); OpenAI; TechInformed.