JMIR Publications

ABSTRACT

As generative artificial intelligence (AI), particularly large language model (LLM)-based tools, is increasingly integrated into diagnosis, triage, decision support, and treatment planning, it offers potential gains in efficiency and information access. However, real-world deployment also introduces important risks, including hallucinations, miscalibrated confidence, automation bias, and increased verification burden on physicians. This burden may divert attention from independent clinical reasoning, contribute to deskilling, and increase vulnerability when models fail silently or perform poorly in unfamiliar clinical contexts. Existing AI governance frameworks emphasize data quality, transparency, accountability, and ethical deployment, but give less attention to physician-facing resilience, defined here as the capacity to sustain independent and safe clinical judgment when collaborating with generative AI. In this viewpoint, we propose a multilevel governance framework organized around three coordinated domains: cognitive workload shaping, clinical authority governance and allocation, and organizational safety governance and accountability. Together, these domains aim to reduce verification burden, preserve physician decisional authority, and align institutional oversight with safe and context-sensitive AI use. The framework includes mechanisms such as risk-sensitive verification triggers, bounded delegation, structured interprofessional review, and organizational monitoring to support safe clinical integration while minimizing avoidable workflow disruption. Rather than presenting these mechanisms as empirically validated solutions, this article offers a structured governance proposal to guide the safer integration of generative AI into clinical care and to inform future evaluation across specialties, workflows, and institutional settings.