Building Trust in Healthcare AI: An Actionable Governance Playbook From a Decade of Implementation Failures (Preprint)

UNSTRUCTURED Healthcare artificial intelligence (AI) governance is rich in principles and increasingly rich in institutional guidance, yet still poor in verifiable operational detail. Consensus resources from the Coalition for Health AI (CHAI) and the Joint Commission now tell health systems which structures and processes to stand up; what remains scarce is failure-derived specificity about who owns each obligation and what evidence proves it was met. This viewpoint translates six documented implementation failures, IBM Watson for Oncology, the Epic Sepsis Model, a racially biased care-allocation algorithm, the field failure of diabetic retinopathy screening, dermatology models that underperform on dark skin, and chatbot-related deaths of vulnerable users, into instruments institutions can apply now, organized by AI lifecycle stage: a five-question guide for selecting a human oversight model, and a lifecycle governance checklist assigning every obligation a named owner and a verifiable evidence artifact, from contractual performance floors and subgroup-disaggregated reporting at procurement to alert burden budgets, rehearsed deactivation, and auditable guardrails for patient-facing conversational systems. The language of trust has misdirected governance toward how AI systems feel rather than how they perform. The goal is not to abolish trust but to calibrate it: reliable systems embedded in accountable institutions, not anthropomorphized algorithms, are what deserve clinical reliance.