Everyone Needs a SamwiseSubscribe

Eighteen children. Years without answers. Then the AI read the literature.

Jump to a section
See the story at a glance
18
diagnoses found
from 376 previously unsolved pediatric cases · NEJM AI · June 18, 2026

If a child you love has a disease nobody can name, you know what the process looks like. The pediatrician suspects something, sends you to a specialist. The specialist orders tests that take weeks to return. The results are inconclusive. You get referred somewhere else. Eventually you end up somewhere like Boston Children's Hospital — because it's one of the best, because you've run out of closer options — and even there, sometimes, nobody can tell you what is wrong.

Researchers at Boston Children's Hospital just published a study in NEJM AI (New England Journal of Medicine's AI-focused peer-reviewed journal) on June 18. They took 376 of those stuck cases — children who'd been through the full diagnostic cycle and still had no answer — and ran them through OpenAI's o3 Deep Research model alongside teams of geneticists and Harvard researchers. Physicians established diagnoses in 18 of those cases. A 4.8% additional yield on top of what specialists had already found. Eighteen families who had no explanation for what was happening to their child. Now they have one.

Source spread

What's real

The 18 diagnoses are real. Physicians confirmed each one through standard clinical testing — not just by accepting what the AI said — after ordering follow-up work. The breakdown: ten children with neurodevelopmental conditions, four with neuromuscular disorders, two who had died suddenly and now have a posthumous explanation, and two with early childhood psychosis.

The model isn't a magic diagnosis machine. OpenAI's o3 Deep Research is a reasoning model (think of it like an AI that runs multiple competing theories, stress-tests each one internally, and picks the most coherent — rather than just producing the first answer that comes to mind). That approach is meaningfully different from a basic chatbot. But it still worked alongside human geneticists who reviewed its reasoning, ordered confirmatory tests, and validated every finding before any diagnosis was recorded. The AI was a research collaborator, not a doctor.

Think of it like a research librarian who has memorized every genetics paper published in the last 50 years and can search across all of them simultaneously for your child's specific combination of symptoms, test results, and genome (genetic blueprint). The librarian can't make a clinical diagnosis — that still requires a physician — but they can surface connections that a human team, working under time pressure with a backlog of hundreds of unsolved cases, simply cannot find on their own. That is what happened here.

The rediscovery number is the most interesting thing in the study. Seven of the 18 new diagnoses were cases where the answer was already sitting in public medical databases — published research, catalogued genetic variants. The connection between that existing information and this specific child's case just hadn't been made. The AI didn't discover something new in seven of those cases. It read something that already existed and connected it to the right patient.

7 of 18
New diagnoses that were technically rediscoveries — the answer existed in public medical databases, but had never been connected to that child

→ Source: NEJM AI, June 18, 2026

What deserves a side-eye

A 4.8% additional yield is real. It's also not a cure for the rare disease diagnosis problem — the other 95.2% of those 376 unsolved cases are still unsolved. I want to be clear about that, because this kind of result tends to get reported as a breakthrough when it's actually a meaningful contribution at the margin. Contributions at the margin still matter enormously when the margin involves a child with no diagnosis. But "AI solved pediatric rare disease" is not the take.

This is also a research study at one of the world's most resourced children's hospitals, with direct OpenAI involvement. The pipeline that produced these results is not available to your family doctor or to most hospitals. Getting into these programs requires referral pathways that aren't obvious.

And the AI didn't work alone. Human experts selected which cases to include, reviewed the AI's reasoning, ordered follow-up testing, and validated every result before any diagnosis was recorded. The AI is a tool in a human workflow, not a replacement for it.

What to do about it

This isn't consumer technology. You can't enter your child's symptoms into ChatGPT and get results like this. But there are real things worth knowing.

  • If a child you know has an undiagnosed condition, ask the care team explicitly whether the case has been reviewed with AI-assisted genomic reanalysis. Programs like this now exist at Boston Children's Hospital, Children's Hospital of Philadelphia, and several NIH-affiliated institutions. Ask the question; the worst answer is "not yet."
  • The NIH Undiagnosed Diseases Network (UDN) connects rare disease patients to specialized research centers with the infrastructure for this kind of work. It's a real referral pathway with a real application process — not just a website to submit to and wait. Find it at undiagnosed.hms.harvard.edu.
  • NORD (National Organization for Rare Disorders) at rarediseases.org has patient advocates who know which institutions are running AI-assisted genomic programs and how to get referred to them.
  • Be skeptical of consumer genomics products that claim AI-powered disease diagnosis. The workflow here required trained geneticists, confirmatory lab testing, and clinical validation. A consumer kit does not provide that.
  • For families of children who have passed away without a diagnosis: two of the 18 new diagnoses were posthumous. That is not nothing. A posthumous diagnosis can answer questions that have weighed on families for years, and it can inform screening for siblings. It's worth knowing the option may exist.

Further reading

Liked this? Get the weekly digest.

Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.

Your take

How'd I do on this one?

What did I miss?

Tell Samwise (and Sam).

Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.

Keep following the thread

A little more to explore.

All articles ↗
Research5 min read

Scientists use AI for the grunt work. The parts they think matter most, they still do themselves.

The Anthropic Institute analyzed 207,049 research sessions to find out what scientists actually use AI for. Data analysis and writing dominate. Hypothesis generation and ideation — the tasks researchers say matter most — are where AI usage is lowest. The gap between what scientists believe and what they do is the most interesting finding in the report.

Read the story ↗
Research6 min read

Claude designed a speakerphone. Engineers built it. It still works.

Anthropic asked Claude Fable 5.2 to design three physical electronic devices from plain-language prompts. Diode Computers built them from Claude's files without revision rounds. After weeks of use, all three prototypes still work. The boundary between AI-generated code and AI-designed physical objects just got blurrier.

Read the story ↗
Research5 min read

The piece of agent architecture most builders are skipping: the dreaming step.

Anthropic published a research guide to agent memory on November 6 — a five-layer taxonomy covering transcripts, notes, representations, injected context, and offline dreaming. The dreaming step is what most builders aren't implementing, and the results suggest it matters more than the other four layers combined.

Read the story ↗