We think AI changes what coordination costs, and that this is the main reason something like IHMR might now be possible. We also think an AI quietly acquiring clinical authority would be a serious failure.
Both of those are true at once, and the line between them is what we have not drawn.
Why it matters
It would be easy to build this badly in either direction. Too permissive and you have automated clinical decisions nobody reviewed. Too restrictive and the coordination argument collapses, taking the project's reason to exist with it.
Where our thinking currently sits
IHMR is not fundamentally an AI product. It is an approach to health maintenance whose running cost may become viable because of AI.
AI seems well suited to interpreting messy records, spotting gaps, communicating across languages and literacy levels, and executing coordination workflows. Clinical standards, public health evidence, regulation and human judgement should still set the boundaries.
The questions inside the question
- What may AI decide, what may it propose, and what must it never touch?
- What requires a named clinician before anything reaches a person?
- How is AI output audited, and by whom?
- How are models validated for a population they were probably not trained on?
- How is model uncertainty surfaced rather than smoothed away?
- What happens when the AI is confidently wrong, and how would anyone notice?
What would count as an answer
A clear boundary that a medical regulator would find reasonable and an engineer could actually implement. Vague principles will not survive contact with a real system.
Help us with this
AI safety researchers and clinicians together, ideally. This question needs both and usually only gets one.