The Clinical AI Safety Gap from Correct Answers to Safer Decisions

High accuracy alone does not establish that language-model assistance improves clinical safety. Drawing on 31 selected publications, this critical Perspective outlines an error-pathway ledger connecting incorrect outputs to display, confirmed viewing, clinician recognition, final decisions, and potential harm. It specifies distinct descriptive indicators and a randomized comparison for evaluating unsafe decisions. The framework provides an operational research design for prospective validation against established clinical evaluation approaches.