Health AI is discussed either as an imminent replacement for clinicians or as an expensive autocomplete. Neither is a useful description of where the technology actually is. Here is a more honest inventory.
What it genuinely does well
- Turning documents into structure. Reading a page that was laid out for human eyes and producing the values, the units and the ranges as data is reliable now, in a way it was not a few years ago.
- Explaining in plain language, at scale. Rewriting a clinical sentence for a non-clinical reader, in any of a very large number of languages, at a cost per document that makes it viable for every patient rather than a few.
- Drafting the repetitive writing. Referral letters, certificates, summaries. Work with a fixed shape, high volume and low ambiguity.
- Holding the thread. Remembering what was ordered, what came back, what is outstanding and what was said last time.
What it is still bad at
- Knowing what it does not know. Calibration remains the central unsolved problem.
- Anything requiring examination. A photograph of a page is not a patient.
- Judgement under genuine uncertainty. Which is a large share of clinical medicine.
The pattern in that list
Everything in the first group is a transformation: take something that exists and render it in another form. Everything in the second requires producing a conclusion that is not present in the inputs.
That is the line worth holding in mind when any health AI is described to you. Transformation is mature. Conclusion is not, and the products that blur the two are the ones to be careful of.
Where the value actually accrues
Not in replacing the expensive judgement, which is not close. In removing the hours of unskilled work wrapped around it — the reading, the transcribing, the chasing, the explaining — so the judgement is the only thing the expensive person spends their day doing.
This article explains how our technology works. It is health information, not medical advice.