Clinical AI oversight review finds human error interception cannot be assumed
Published in Artificial Intelligence, Journals, Tools.
Published November 1, 2026.
A systematic review published in npj Digital Medicine on October 23 examines whether clinicians detect and resist incorrect automated advice.
Zeng and colleagues identified 46 reports representing 41 independent cohort clusters, including 40 core cohorts. Across three adjusted cohorts, clinician correctness was substantially higher when automated advice was correct rather than incorrect. This association does not by itself quantify the benefit of AI over unaided care.
In four highly heterogeneous real-world studies of computerized ECG interpretation, an estimated 26.6% of erroneous outputs were not recognized or corrected. The uncertainty was substantial: the 95% confidence interval ranged from 6.4% to 65.6%. That pooled estimate should not be treated as a universal error rate for modern AI or chatbots.
Contemporary AI and language-model studies also documented harmful reliance but were too diverse to pool. Patient-level adverse events were reported in only one retrospective cohort. The review supports testing oversight as a measurable ability to intercept errors. The authors report no study funding or competing interests.
Source: Zeng et al., npj Digital Medicine, October 23, 2026.