Healthcare AI only works when it earns clinical trust

Every impressive healthcare AI demo I have seen shares the same weak point: it answers confidently and cannot show its work. In a clinic, an answer without a source is not an answer — it is a liability.

The pattern that actually survives contact with practitioners is narrow and grounded. Give the model a defined corpus: the clinic's own protocols, the manufacturer's instructions, the payer's policy document. Then require a citation with every claim, and require the model to say plainly when the answer is not in the corpus.

This feels less magical, and it is far more useful. A dental assistant that reliably retrieves the correct sterilisation protocol with the paragraph attached will be used every day. A general chatbot that occasionally invents a drug interaction will be switched off within a week — correctly.

The design work, then, is mostly boundary work: what is in scope, what the refusal looks like, who reviews the corpus, and how errors get reported. Model choice is the easy part.