Abstract
Background
Artificial intelligence (AI) is increasingly integrated into healthcare for diagnosis, risk prediction, and clinical decision support. However, ensuring its safe use requires more than technical performance; it depends on governance processes that coordinate technology, clinical practice, and organisational systems throughout the AI lifecycle. This study aimed to explore how ethical principles underpinning AI safety are operationalised in clinical practice and to identify the governance conditions required to translate these principles into safe AI implementation.Methods
A qualitative study (QuAS-AI) was conducted using semi-structured interviews with 16 experts involved in clinical AI implementation, including academics, industry professionals, and practitioners. Participants were recruited through purposive and snowball sampling to capture expertise across diverse clinical AI implementation contexts rather than to represent any single clinical specialty.The trustworthiness of the findings was strengthened through independent coding, peer debriefing, member checking, and conceptual validation.Results
Three interrelated themes emerged. First, performance, risk, and bias were conceptualised as dynamic processes requiring continuous monitoring across clinical contexts. Second, transparency and explainability were viewed as complementary mechanisms supporting clinical interpretation rather than purely technical disclosure.Third, accountability was framed as a multi-actor governance structure involving clinicians, developers, and healthcare organisations, which integrates traceability, logging, data governance, and intervention capacity. Findings suggest that AI safety depends on ongoing monitoring, feedback integration, and institutional oversight rather than one-time validation.Conclusion
AI safety in healthcare should be understood as a socio-technical governance process rather than a fixed technical property. A High-Reliability Organisation approach may provide a useful framework for early risk detection and continuous oversight. Safe implementation requires integrated accountability and data governance mechanisms that support clinician oversight and institutional responsibility throughout the AI lifecycle. Future research should examine how pre-implementation governance activities relate to post-deployment mechanisms and should include bedside clinicians and patients.