When clinical AI fails, review the case, not just the model

What the source reports

The arXiv abstract says current clinical AI safety tools do not fit single AI errors and near misses. Model checks can find changes. Patient safety reports can log harm. Neither shows how risk forms across AI systems, doctors, job steps, and local controls. The proposed AI Morbidity and Mortality framework is blameless and case based. It uses standard intake and saved proof. It also uses investigator review, tool-in-loop attribution, and tracking for fixes. Events are sorted by Trigger, Mechanism, Clinical Pathway, and Corrective Action. The authors show it on five outpatient medication and decision-support cases. Two clinician reviewers agreed across all 20 axis-level classifications. They say future testing is still needed.