The dominant idea in AI for the last several years has been generality: one system that does everything, improving across every subject at once. It is a good idea, and for most work it is the right one. Health documents are where it stops being right.
Breadth and reliability pull against each other
A system built to answer anything is optimised to produce an answer. That optimisation is invisible until the subject is one where the honest response is frequently I cannot tell you that from this document. In health, that response is not a failure — it is often the only correct output.
Building for health means choosing reliability over coverage, and accepting that the system will decline questions a general assistant would happily attempt.
What we built the model to be good at
- Reading a document it has never seen before. Not a tidy data feed — a photograph of a page, taken at an angle, in whatever layout that particular laboratory uses.
- Staying attached to what is printed. Values, units and the range on the page are treated as the source of truth, not as a prompt for recall.
- Answering in the language of the question. Including the way people actually speak, which is often more than one language in the same sentence.
- Refusing cleanly. Declining is a designed behaviour with its own wording, not an error state.
What we deliberately did not build it to do
It does not name conditions. It does not recommend treatment. It does not take a person’s values and return a diagnosis, because a system that does that is a regulated medical device and should be held to that standard. We give people information about what is written on their document, and we say who to speak to and how soon.
That boundary is not a stage we are passing through on the way to something less careful. It is the product.
This article explains how our technology works. It is health information, not medical advice.