If you ask most people what makes a health AI good, they will describe the quality of its answers. In our experience the harder and more consequential problem is the opposite one: making the system stay quiet when it should.
Why silence is difficult to engineer
Language models are trained on text in which questions are followed by answers. The pattern is overwhelming, and it pulls every system towards responding. Getting one to stop — reliably, in the right situations, without becoming uselessly cautious everywhere else — takes deliberate work against that grain.
The failure mode on either side is real. Too eager and it invents. Too cautious and it refuses ordinary questions, which teaches people it is useless and they stop asking.
The situations where we stay quiet
- The document does not say. If a value is not on the page, we do not supply one from memory.
- The question is a diagnosis in disguise. “What do I have?” has no safe answer from a document, however it is phrased.
- The page is not what it appears to be. If the upload is not a medical document, we say so rather than reading something into it.
- The answer would need a number we cannot source. If we cannot attribute it, we do not state it.
Saying nothing is not the same as being unhelpful
Declining is only acceptable if it comes with somewhere to go. When we cannot answer, we say what is worth asking and who is worth asking it of. The person leaves with a better question rather than a confident answer that was never earned.
A system that says “I do not know” is giving you real information about its own confidence. One that never says it is giving you none.
This article explains how our technology works. It is health information, not medical advice.