An OpenAI reasoning model just cracked 18 previously unsolvable rare disease cases in children — and that's not a metaphor for "helped a bit", it means kids who had no diagnosis now have one.
Why Rare Disease Diagnosis Is So Hard (And Why AI Changes That)
Rare genetic diseases are medicine's hardest puzzle. There are roughly 7,000 of them, most affecting fewer than 1 in 10,000 people, which means no single doctor can hold the pattern-recognition expertise for all of them in their head.
That's exactly where AI-assisted rare disease diagnosis earns its keep. Researchers fed previously unsolved paediatric cases into an OpenAI reasoning model — the kind built to think in steps, not just retrieve facts — and it identified diagnoses in 18 cases that had stumped human clinicians.
What the AI Actually Did Differently
This wasn't a chatbot guessing symptoms. Reasoning models work by chaining logical steps, weighing evidence, and interrogating their own conclusions — closer to a specialist working through a differential diagnosis than a search engine returning results.
The model could cross-reference phenotypic clues (physical symptoms, lab values, family history) against the vast landscape of rare genetic conditions simultaneously, something no individual physician has the bandwidth to do in a clinical setting. The result: 18 new diagnoses from a pool of "previously unsolved" cases. For those families, this is life-changing.
If you want to understand the architecture that makes this kind of multi-step reasoning possible, our How Neural Networks Really Work course breaks down exactly how these models build and evaluate chains of inference.
What This Means for Learners
This story is a masterclass in understanding which AI tool fits which problem. A standard language model summarises. A reasoning model deliberates. Knowing the difference is a core AI literacy skill that separates people who use AI effectively from those who don't.
Healthcare is also becoming one of the highest-stakes arenas for AI deployment — which means understanding how to evaluate AI outputs critically, and where human oversight is non-negotiable, is increasingly valuable. Our When AI Goes Rogue course covers exactly how to think about AI reliability in high-stakes contexts like this one.
The practical takeaway: next time you're choosing an AI tool for a complex problem, ask yourself — do I need a model that retrieves, or one that reasons? That question alone will make you a sharper AI user.
