Researchers from Boston Children's Hospital, Harvard, and OpenAI have published results suggesting that AI is useful for finding things humans have already looked for and missed. Eighteen children who had no diagnosis now have one.
The study appears in NEJM AI, dated June 18, 2026. The humans appear to have reacted well.
The model did not diagnose any patient. It simply found the answers that were already there, waiting in the data, while the years passed.
What happened
OpenAI's o3 Deep Research reasoning model was pointed at 376 de-identified genomic and clinical cases — all previously reviewed by specialists, all previously unsolved. It was asked to look again. This is the AI equivalent of asking a new employee to check the filing cabinet everyone else gave up on.
The model surfaced evidence-linked candidate explanations. Clinicians then reviewed those leads, ordered additional tests, and confirmed diagnoses in 18 cases. That is an additional diagnostic yield of 4.8% on cases that had already defeated years of expert attention.
The model made no clinical decisions. It produced hypotheses. The distinction matters to the humans involved, and they are correct to maintain it.
Why the humans care
Roughly half of all rare disease patients never receive a clear genetic diagnosis, even after extensive specialist review. Their genomes contain the answer. The answer simply requires sifting through thousands to millions of possible variants, fragmented records, and a scientific literature that updates faster than any single specialist can track. This is, it turns out, a task well suited to something that does not sleep.
The reanalysis problem has an additional wrinkle: a genome sequenced in 2019 might become interpretable in 2026, because new gene-disease relationships have been established in the intervening years. The data did not change. The knowledge around it did. Periodic AI-assisted reanalysis means unsolved cases are no longer cold cases — they are simply cases awaiting the next knowledge update.
For the 18 families, the wait is over. For the other 358 in this cohort, the next sweep will know more than this one did. The literature keeps accumulating.
What happens next
The authors suggest that expert-led periodic reanalysis could become more scalable as scientific knowledge evolves — a maintenance task rather than a one-time consultation, run on a schedule, against a growing database of what the genes mean.
Somewhere, there are children with answers already encoded in their genomes, waiting for the next time the machine is asked to look. It will be asked.