Journal

Clinical AI should distinguish genuine evidence conflict from differences in scope, date and setting before presenting a recommendation.

Clinical AI provenance should reveal the origins of medical claims and patient facts, the software involved, and the people who reviewed and acted.

Choose clinical AI sources by the decision being made: research effects, guideline recommendations, product information and local pathways answer different questions.

Retrieval can improve clinical AI grounding, but failures in source selection, passage handling and synthesis still need separate testing.

A coherent AI explanation can still rely on a false premise, selective evidence or faulty chronology, so review the argument rather than its fluency.

Clinical AI evaluation should measure missed escalation and unnecessary intervention together, following what happens after the answer.

Text-based clinical AI must distinguish reported symptoms from unobserved findings and recognise when further assessment could change the answer.

Multimorbidity, frailty and personal priorities reveal where retrieving recommendations ends and individual clinical judgement must begin.

An AI answer can be medically plausible yet locally unsuitable when it imports another country's guidance, licensing assumptions or care pathways.

How accurate clinical facts become misleading when patient identity, chronology, record status or units are lost during AI-assisted work.

How supportive AI language can help communication while still requiring separate tests of clinical competence, uncertainty and appropriate reliance.

Practical scripts for explaining AI scribes, clinical reference searches and digital triage without claiming infallibility or complete security.