Tom Bisson, Henriette Voelker, Sanddhya Jayabalan et al.
Large language model (LLM)-based AI agents are increasingly capable of complex clinical reasoning and may soon participate in medical decision-making with limited or no real-time human oversight. This shift raises fundamental questions about how the core principles of medical ethics (i.e., beneficence, nonmaleficence, autonomy, and justice) can be upheld when the clinical responsibility extends to autonomous systems. Here we propose an ethics-by-design framework for medical AI agents comprising six practical interventions: auditable ethical reasoning modules, explicit human override conditions, structured patient preference profiles, AI-specific ethics oversight tools, global benchmarking repositories for ethical scenarios, and regulatory sandboxes for real-world evaluation. Together, these mechanisms aim to operationalize ethical governance for emerging clinical AI agents. https://github.com/BissonTom/Ethical-Governance-of-Medical-AI-Agents