Clinical Oversight Is Not a Safety Net Unless It Catches the Error
A systematic review finds a large gap between performance with correct and incorrect automated advice—and warns that merely placing a clinician in the loop does not establish safety.
Direct answer
Does human review make erroneous clinical AI safe?
Not automatically. Across the evidence reviewed, clinicians were much more likely to be correct when automated advice was correct than when it was wrong. The authors argue that oversight should be measured as an error-interception capability rather than assumed from the presence of a human reviewer.
What the review found
Di Zeng, Jianyong Lei and Zhihui Li reviewed 46 reports representing 41 independent cohort clusters, including 40 core cohorts. In three adjusted cohorts, clinician correctness was substantially higher with correct than incorrect automated advice, with a pooled odds ratio of 10.19 and a 95% confidence interval from 7.43 to 13.98. The peer-reviewed article was published early by npj Digital Medicine on 23 October 2026 and may still receive copyediting before the final version of record.
Four heterogeneous real-world studies of computerized ECG interpretation found that an estimated 26.6% of erroneous interpretations were not corrected or recognized. That pooled figure is uncertain: its 95% confidence interval ran from 6.4% to 65.6%, and heterogeneity was very high at I² = 95.8%. Contemporary large-language-model studies also documented harmful reliance but were too varied for pooling.
Why 'human in the loop' is an incomplete control
A human can only intercept an error if the interface exposes enough evidence, the reviewer has relevant expertise and time, and organizational incentives permit disagreement. Automation bias, workload, anchoring and confidence displays can all change that behavior. A nominal approval click is therefore not equivalent to independent verification.
The review makes a useful distinction between overall accuracy and conditional resilience. A system can look strong on average while creating a dangerous failure mode if wrong advice systematically pulls users away from correct judgments. Evaluation should deliberately inject plausible errors and measure whether the combined human-machine process recovers.
What remains unresolved
The evidence base spans different tasks, interfaces, technologies and study designs. Only one retrospective cohort reported patient-level adverse events, so the review cannot estimate a general rate of clinical harm. The wide intervals in real-world ECG studies should not be flattened into a single universal percentage.
Prospective validation should specify the error classes, reviewer population, decision stakes, time pressure and escalation path. It should test not only whether clinicians notice a wrong suggestion but whether the final action is corrected and whether the control remains effective after routine use changes attention and trust.
Primary sources and version boundary
- Zeng, Lei and Li, npj Digital Medicine, peer-reviewed early publication, 23 October 2026
- Permanent DOI
Published figures and claims are reported from the linked source version. The featured researchers and organisations did not commission or endorse this coverage. This article is independent Discovery News reporting and is not a Haivian audit result.
Independent validation
Need a computational claim examined before publication?
Haivian offers scoped research audits with explicit evidence, limitations and accountable human sign-off. Payment does not purchase a positive result or publication.
Explore research audits