Eighteen families got an answer. The other 358 are the story too.

A peer-reviewed study published today in NEJM AI reports that OpenAI's o3 reasoning model, run across 376 previously unsolved pediatric rare-disease cases at Boston Children's Hospital, helped clinicians reach 18 new diagnoses after expert review. That is a real result for 18 families, and the most carefully evidenced AI-in-medicine claim I have read this year. It is also a result that needs reading in two directions at once: what it proves about the tool, and what it reveals about the system the tool is being used inside.

What the study actually shows. Researchers at Boston Children's Manton Center for Orphan Disease Research, working with Harvard Medical School and OpenAI, reanalysed 376 cases that had already failed expert human review. The o3 model integrated clinical notes, phenotype descriptions, and filtered gene lists, and proposed candidate diagnoses. Physicians then reviewed every suggestion before any diagnosis was recorded. Eighteen were confirmed, across neurodevelopmental disorders, neuromuscular disorders, sudden unexplained deaths, and early childhood psychosis.¹

A 4.8% additional yield on a pre-screened, already-failed population is not a small number. These are families who, by the definition of the cohort, had been told by some of the best rare-disease specialists in the world that there was no answer to give them yet. For a rare-disease family in the United States, the average diagnostic odyssey runs 4.8 years and 7.3 physicians.² Eighteen of those odysseys ended this year because a reasoning model surfaced a candidate a human team had missed and a human team then confirmed it. I am not going to pretend that is anything other than good.

Why this study is unusual. Most AI-in-healthcare claims arrive as press releases without a denominator, without a numerator, and without peer review. This one has all three, plus a mandatory physician confirmation layer, plus a published methodology in a journal run by the NEJM Group. The deployment pattern, defined population, defined outcome, human-in-the-loop, peer-reviewed write-up, is what makes the result credible. The absence of this pattern almost everywhere else in the AI-in-medicine literature is, on its own, worth naming.

So far, so good. Here is where I want to widen the frame.

Who was in the room, and who was not. The 376 cases in the study had already reached the Manton Center — one of the world's leading rare-disease programs, embedded in a top-five US children's hospital, attached to Harvard. Getting a case into that 376 is itself a substantial resource advantage. The roughly 30 million Americans with a rare disease, and the global population estimated at 300 million or more, are overwhelmingly not seen by anyone like Manton.² OpenAI's framing — that this study shows reasoning models can "help democratise access to specialist knowledge" — is aspirational. It is not descriptive of how the technology is currently being deployed. Right now it has been deployed at the place that already has the most specialist access in the country.

I want to take the democratisation claim seriously, because the underlying capability is real and the gap it could close is enormous. But democratisation is a verb that requires a subject. Who is going to fund the deployment of this pattern at the community hospital in a mid-sized US city, where most rare-disease patients first present and most diagnostic odysseys begin? Who pays for the physician confirmation layer in a health system that does not have a Manton Center to staff it? The study does not answer those questions because that is not what the study is. But the coverage so far has treated the access problem as already solved by the existence of the capability, and it is not.

The seven-million-dollar number deserves its own paragraph. Boston Children's separately reports that its broader AI integration has saved 60,000 work hours and redeployed $7 million in labour costs. That figure is the hospital's operational self-report, not a finding of the NEJM study, and there is no independent audit alongside it.¹ I have read enough of these productivity claims, from PwC's Jobs Barometer through every major consulting house, to know that "redeployed" is the load-bearing word and the one least often defined. Redeployed to whom? Hours saved by which workers, doing what afterwards? Cost avoided by not hiring, or cost recovered from existing payroll? These are not rhetorical questions. The same hospital can produce a genuine clinical gain for 18 families and a labour claim that does not survive audit; both can be true. The clinical result does not vouch for the operational one.

Consent is the part nobody is writing about. De-identified genomic data was sent to a closed commercial reasoning model. De-identification of rare-disease genomic profiles carries a higher residual re-identification risk than for common conditions, because the patient populations are by definition small and the phenotypes distinctive. There is no public documentation, in the study or in the coverage I have read, of what patients and families were told about commercial AI processing of their case records. I am not claiming anything improper happened. I am claiming that an architecture where a child's de-identified genome and clinical narrative travels to a vendor API is a consent question, and that the absence of any discussion of that question in the public materials is itself a finding.

Where this leaves me. I think the result is real. I think 4.8% additional yield on cases the world's best human reviewers had given up on is the kind of evidence that warrants the technology being taken seriously. I also think that if the next 24 months of coverage treat this study as a near-solution to rare-disease diagnosis, rather than as proof that one pattern works at one site for one cohort, the actual systemic problem — that most rare-disease patients will never reach a Manton Center to be in the 376 in the first place — will get harder to see, not easier. The capability is the easy part now. The distribution is the hard part, and the distribution is what determines whether 18 families this year becomes 18,000 families a decade from now, or whether it stays 18.

Footnotes

¶