AMA issues governance framework for clinical-role AI

Key takeaways

• The AMA’s framework requires AI systems that take on clinical functions to have their role defined, performance validated, and safeguards documented.
• Clinical authority for AI must be earned by evidence and can change as performance data, risk levels, and technology evolve.
• The guidance will shape AMA work on liability, licensure and regulation, directing collaboration with clinicians, developers and policymakers.

The American Medical Association (AMA) has released a governance framework that places a evidentiary gate on any artificial intelligence (AI) solution that performs clinical functions. The central premise is simple: an AI system’s clinical authority must be earned through rigorous evidence and retained only while its performance justifies it. This principle drives every element of the AMA’s approach, from role definition to accountability mechanisms.

Defining the clinical role of AI

The framework begins by obligating developers and health systems to articulate precisely what clinical decision-making the AI will undertake. Whether the tool provides diagnostic suggestions, triage recommendations, or treatment planning, the intended function must be documented before deployment. This explicit role-mapping prevents “black box” adoption where clinicians cannot determine the scope of AI influence.

Validating performance and safeguards

Evidence of safety and efficacy is the next requirement. The AMA calls for prospective validation studies, real-world performance monitoring, and transparent reporting of error rates. Safeguards such as human-in-the-loop checks, fail-over protocols, and periodic re-training must be baked into the system design. The framework emphasizes that validation is not a one-off event; ongoing data collection must feed back into oversight as the AI adapts to new datasets or clinical contexts.

Setting limits on authority and accountability

Once a role and performance baseline are established, the AMA mandates limits on the AI’s decision-making power. For higher-risk functions—such as prescribing medication or determining surgical candidacy—the framework requires that the AI’s output be reviewed and approved by a qualified clinician. Accountability structures are also prescribed: organizations must track AI-driven outcomes, attribute responsibility for adverse events, and disclose these metrics to regulators and patients where appropriate.

Guiding liability, licensure and regulation

The AMA frames its governance work as a blueprint for broader policy discussions. By anchoring clinical authority in evidence, the association aims to inform liability standards, clarify licensure requirements for AI-enabled providers, and shape regulatory pathways. The organization plans to work with physicians, technology developers, and policymakers to translate the framework into actionable rules that can be adopted across the health system.

The urgency of this effort is underscored by AMA CEO John Whyte’s statement to Fierce Healthcare: “AI is advancing toward taking on increasingly consequential clinical functions, which is raising urgent questions about clinical authority, oversight and accountability.” He added that the AMA cannot afford to wait for trial-and-error learning in patient care, and that the new framework sets an “ambitious agenda” to address liability, licensure and regulation now.

The governance model builds on earlier AMA initiatives, including a May-issued framework protecting physicians from AI-generated deepfakes and an April appeal for stricter safeguards on AI chatbots used in mental health. Together, these policies reflect a broader strategy to safeguard clinician authority and patient trust as AI becomes more pervasive in care delivery.

Next steps for health leaders 1. Map AI use cases – Identify every AI tool that performs or influences clinical decisions and document its intended role. *Trade-off*: Requires time and cross-department coordination but prevents hidden scope creep. 2. Commission evidence reviews – Conduct or acquire validation studies that meet the framework’s standards, and embed continuous performance monitoring. *Trade-off*: Upfront cost and effort, offset by reduced risk of adverse outcomes and liability. 3. Implement oversight controls – Establish human-in-the-loop checkpoints proportional to the AI’s risk level, and create audit trails for AI-generated recommendations. *Trade-off*: May slow workflows, yet ensures accountability and regulatory compliance. 4. Define accountability pathways – Assign clear responsibility for AI-related incidents, integrate reporting into existing quality-improvement structures, and communicate policies to staff and patients. *Trade-off*: Adds administrative burden but clarifies liability and supports trust.

By following this ordered approach, medical organizations can align with the AMA’s expectations, mitigate legal exposure, and ensure that AI augments rather than supplants clinician expertise.