What the research found
Researchers at West China Hospital, Sichuan University pooled 46 reports covering 41 independent groups of clinicians to measure how often people catch errors made by clinical AI and other computerised decision support. In the three studies that adjusted for confounders, clinicians were about 10 times more likely to be right when the advice was correct than when it was wrong, and across four real-world ECG studies 26.6% of erroneous computer interpretations went uncorrected, ranging from about 1 in 16 to 2 in 3 between studies. The review covers older, non-AI decision support as well as newer AI tools, and only one included study measured actual patient harm.
A clinician in the loop is a safeguard only if they reliably spot when the tool is wrong, so providers should measure how often their staff override bad advice rather than assume they do.