MedicineFirst.org

Global, equitable AI governance

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights
• Interoperability of governance approaches

These four priorities converge on a single unaddressed governance gap: AI systems operating at the point of individual patient care, Medical AI, are not governed as medical devices despite functioning as medical instruments. Safe, secure and trustworthy AI is our primary priority because the empirical evidence demands it. Published research demonstrates that leading medical AI systems fail specialty board certification standards in complex clinical reasoning, with subspecialty accuracy approaching chance levels. Separate research confirms that access to source literature does not confer the ability to reason through clinical questions derived from that literature. A system deployed to emergency departments in New Zealand was successfully manipulated via adversarial prompt to output dangerous guidance. These are not theoretical risks. They are documented failures in deployed systems used by hundreds of thousands of clinicians. Transparency, accountability, and human oversight is urgent because the prevailing human-in-the-loop model inverts the accountability architecture. The AI acts first. The clinician reacts. The clinician bears full liability. The AI company carries a disclaimer. The January 2026 FDA guidance codified this inversion by exempting clinical AI from premarket review. Authority exercised only at the moment of approval is not authority. It is liability. Protection and promotion of human rights is inseparable from Medical AI governance. Retrospective governance guarantees that populations with the least capacity to detect AI failures bear the highest burden of those failures. The right to safe healthcare requires prospective standards, not post-harm detection. Interoperability of governance approaches is essential because medical credentialing infrastructure already operates across jurisdictions. Specialty boards, device classification frameworks, and competency standards exist globally. Extending them to Medical AI requires interoperability - not invention. Our written submission and companion governance protocol detail the framework for all four priorities.

The listed themes do not distinguish between AI systems that operate on populations and AI systems that operate on individual patients. This distinction is the most significant governance gap currently unaddressed by any international framework. Healthcare AI optimizes scheduling, allocates resources, manages supply chains, and processes population-level data. When it fails, systems retrain. Patients are inconvenienced. Error tolerance exists. The retrospective governance model, deploy, monitor, and correct, is adequate for this problem class. Medical AI informs decisions about a specific patient, at a specific moment, with consequences that may be irreversible. When it fails, a human being is harmed. The retrospective model means the detection mechanism for governance failure is patient harm. The first evidence that the system should not have been deployed is that someone was injured or killed. These are not the same problem class. Governing them under a single framework produces a framework adequate for neither. In January 2026, the FDA exempted clinical decision support software from premarket review. No pre-deployment competency standard is required. Published research demonstrates that leading medical AI systems fail specialty board certification standards in complex clinical reasoning despite widespread clinical deployment. The American Medical Association has called for governance architecture that explicitly rejects this approach. Medicine already operates a prospective governance framework for high-consequence individual decisions for both practitioners and their instruments. Boards, licensure, device classification, premarket evaluation. This infrastructure exists. The mandate to extend it to AI systems operating at the point of individual patient care does not. We have submitted a detailed written contribution and companion governance protocol addressing this gap, co-authored with Dr. Enam Satti. We respectfully ask that Medical AI be recognized as a distinct governance category in the thematic framework for Geneva.

The governance gaps identified above are affecting medicine globally in real time, not prospectively. The most significant challenge is the regulatory retreat from premarket oversight. In January 2026, the FDA exempted clinical decision support software from premarket review. The EU removed a key provision targeting clinical AI. The two largest regulatory jurisdictions simultaneously weakened pre-deployment oversight for the highest-consequence AI application in any industry. The result: AI systems that fail specialty board certification standards in complex clinical reasoning face no regulatory barrier to clinical deployment in any jurisdiction, provided they present output for clinician review. This creates an immediate patient safety crisis that is unevenly distributed. The FDA exemption's safety mechanism, clinician review, presumes institutional infrastructure that exists in academic medical centers but not in rural clinics, not in underserved communities, and not in the low- and middle-income countries where Medical AI is most urgently needed. The populations with the least capacity to detect failures bear the highest burden when those failures occur. For physicians specifically, the accountability asymmetry is now formalized. The American Medical Association identifies liability for AI errors as the primary concern preventing physician adoption. Physicians carry malpractice insurance. AI systems carry disclaimers. No governance framework apportions this liability appropriately at the point of care. The most significant opportunity is that the governance infrastructure required to close this gap already exists. Medical specialty boards have administered competency standards for decades. Device classification frameworks operate across jurisdictions. Credentialing bodies have the authority and methodology to evaluate clinical competence. Extending this architecture to AI systems operating in the same problem class requires authorization, not invention. Medicine is the only sector that built prospective governance for high-consequence individual decisions before AI existed. That architecture is ready to be applied. What is missing is the mandate.

The Dialogue's most consequential role is one no other body can play: authorizing the extension of existing governance infrastructure to AI systems operating in problem classes that infrastructure was designed to govern. In medicine, that infrastructure exists. Medical specialty boards administer competency standards across jurisdictions. Device classification frameworks govern instruments from surgical tools to diagnostic imaging systems. Credentialing bodies maintain the authority to restrict or revoke practice when standards are not met. This architecture governs high-consequence individual decisions prospectively — before harm occurs, not after. No single nation can extend this architecture to Medical AI unilaterally. The FDA moved in the opposite direction in January 2026. National regulatory action has proven insufficient. The Dialogue is uniquely positioned to establish the international expectation that AI systems operating at the point of individual patient care meet pre-deployment competency and safety standards equivalent to those required of every other instrument used in clinical decision-making. Specifically, the Dialogue can play three roles no other body currently fills. First, it can formally recognize Medical AI as a distinct governance category, establishing the taxonomic foundation that all subsequent policy requires. Without this distinction, governance frameworks will continue to treat clinical decision support systems and scheduling algorithms as the same category of risk. Second, it can convene medical credentialing bodies and AI governance bodies in the same room. These communities do not currently coordinate. Specialty boards have the methodology. AI governance bodies have the policy mandate. Neither alone can close the gap. Third, it can establish the international expectation that prospective governance, demonstrated competency before deployment, is the appropriate model for AI in individual patient care. That expectation, once established, becomes the floor that national regulators build upon rather than retreat from.

The Dialogue should build upon four categories of existing infrastructure rather than constructing parallel mechanisms. Medical credentialing bodies. Specialty boards worldwide, the American Board of Medical Specialties system, the Royal Colleges, and equivalent bodies across every WHO member state administer validated, standardized competency assessments in every clinical domain where AI is now deployed. These bodies have the methodology, the domain expertise, and the institutional authority to evaluate whether an AI system reasons at the level of the clinical decisions it informs. The Dialogue should engage them directly as partners in developing domain-specific AI competency standards. The World Health Organization. The WHO's Ethics and Governance of AI for Health framework has identified the requirement for physician authority in clinical AI deployment but has not yet provided the technical architecture to fulfill it. The Lead Zero Standard, submitted as a companion governance protocol to our written contribution, provides that operational specification. The Dialogue should direct WHO to incorporate pre-deployment competency standards and clinician-authority architecture into its existing framework. The American Medical Association and national physician organizations. The AMA has independently identified every governance requirement proposed in our submission, pre-deployment validation, independent testing, appropriate liability apportionment, and rejection of the "try-first" mentality. National physician organizations in multiple countries have adopted parallel positions. The Dialogue should recognize this existing physician consensus and build upon it rather than relitigating questions the clinical community has already resolved. AI2030 and multistakeholder governance initiatives. AI2030's Global Fellows program brings together practitioners across AI governance, clinical practice, technical architecture, and health equity. Our submission was developed through this network. The Dialogue should leverage such cross-disciplinary networks to ensure that governance proposals reflect both policy requirements and implementation realities. The added value of the Dialogue is convening authority. The pieces exist. The mandate to assemble them does not.

The most critical structural recommendation: the Dialogue must include domain practitioners alongside policy generalists. In medicine, this means physicians who use clinical AI systems and who bear the consequences when those systems fail. The current multistakeholder model risks producing governance frameworks designed by people who have never held malpractice liability for an AI-assisted clinical decision.Three format recommendations follow from this principle.First, thematic breakout sessions should be organized by problem class, not by technology type. A session on "AI in healthcare" that treats scheduling optimization and clinical decision support as one topic will produce governance adequate for neither. The Healthcare AI / Medical AI distinction should structure the health-related sessions.Second, domain-credentialing bodies should be formally invited as participants, not observers. Medical specialty boards, nursing credentialing bodies, and pharmacy boards possess the institutional methodology to evaluate clinical competence. They are the existing governance infrastructure this Dialogue should activate. Their absence from prior consultations is a structural gap.Third, the Dialogue should adopt a "written submission plus oral intervention" model that weights substantive proposals over procedural statements. The March consultation demonstrated that three-minute interventions can introduce concrete governance frameworks. The July Dialogue should allocate time for proposers of written submissions to present and defend their frameworks, not merely summarize them.Regarding stakeholder contributions specifically: technology developers should be required to disclose pre-deployment testing methodology for any system operating in high-consequence individual decision environments. Physicians and patient advocates should evaluate whether proposed governance frameworks protect the populations they serve. Credentialing bodies should assess whether their existing competency infrastructure can be extended to AI systems. Each stakeholder contributes what they uniquely possess. The Dialogue assembles what no single stakeholder can build alone.

Question 15: How can different stakeholders contribute? Recommendations for format and structure. The most critical structural recommendation: the Dialogue must include domain practitioners alongside policy generalists. In medicine, this means physicians who use clinical AI systems and who bear the consequences when those systems fail. The current multistakeholder model risks producing governance frameworks designed by people who have never held malpractice liability for an AI-assisted clinical decision. Three format recommendations follow from this principle. First, thematic breakout sessions should be organized by problem class, not by technology type. A session on "AI in healthcare" that treats scheduling optimization and clinical decision support as one topic will produce governance adequate for neither. The Healthcare AI / Medical AI distinction should structure the health-related sessions. Second, domain-credentialing bodies should be formally invited as participants, not observers. Medical specialty boards, nursing credentialing bodies, and pharmacy boards possess the institutional methodology to evaluate clinical competence. They are the existing governance infrastructure this Dialogue should activate. Their absence from prior consultations is a structural gap. Third, the Dialogue should adopt a "written submission plus oral intervention" model that weights substantive proposals over procedural statements. The March consultation demonstrated that three-minute interventions can introduce concrete governance frameworks. The July Dialogue should allocate time for proposers of written submissions to present and defend their frameworks, not merely summarize them. Regarding stakeholder contributions specifically: technology developers should be required to disclose pre-deployment testing methodology for any system operating in high-consequence individual decision environments. Physicians and patient advocates should evaluate whether proposed governance frameworks protect the populations they serve. Credentialing bodies should assess whether their existing competency infrastructure can be extended to AI systems. Each stakeholder contributes what they uniquely possess. The Dialogue assembles what no single stakeholder can build alone. 299 words. Question 16: Which voices are currently underrepresented? How could they be included? Three communities are systematically underrepresented in global AI governance discussions, and their absence produces governance frameworks that fail at the point of greatest consequence. Practicing clinicians who use Medical AI daily. The physicians, nurses, and clinical pharmacists who interact with AI-assisted clinical decision support in real patient care are rarely present in governance discussions. Their clinical reality — where AI accuracy drops to chance levels in subspecialty reasoning, where alert fatigue degrades warning systems, where time-critical decisions cannot accommodate "independent review" of AI output — is not reflected in frameworks designed by policy professionals and technology developers. Inclusion requires direct invitation to credentialed clinicians, not merely to professional organizations that represent them at a policy level. Patients and patient advocacy organizations from low- and middle-income countries. The populations most likely to be harmed by ungoverned Medical AI are the least represented in discussions about how to govern it. The solo practitioner in a rural clinic in sub-Saharan Africa who relies on AI output because no specialist is available. The patient who has no recourse when a confidently wrong recommendation produces an irreversible outcome. These are not hypothetical stakeholders. They are the sentinel populations whose harm is the detection mechanism under the current retrospective governance model. Their inclusion requires funded participation, not open invitations that presume resources for travel and attendance. Medical credentialing bodies. Spec

The most effective engagement format for the July Dialogue would invert the standard panel structure: instead of experts presenting to an audience, proposers of written submissions should defend their governance frameworks against structured challenge from domain practitioners. Governance stress-testing sessions. Each proposed governance framework for a high-consequence domain should be subjected to scenario-based challenge by practitioners from that domain. For Medical AI, this means presenting a governance proposal to practicing physicians and asking: does this framework protect your patient when the AI is wrong at 2 AM in a rural clinic with no specialist backup? If the proposers cannot answer that question with specifics, the framework is not ready for the Co-Chairs' summary. This format produces actionable governance rather than consensus language. Cross-domain architecture workshops. Medicine is not the only domain with high-consequence individual decisions. Aviation, nuclear safety, and pharmaceutical manufacturing all operate prospective governance architectures. A workshop that maps the structural parallels — premarket evaluation, credentialing, post-deployment surveillance, liability architecture — across these domains would demonstrate that prospective governance is not a medical exception but a proven model for any domain where error is individual and irreversible. The Dialogue should facilitate this cross-pollination rather than treating each sector in isolation. Practitioner-regulator direct dialogue. A structured session pairing clinicians who use Medical AI with regulators who exempted it from premarket review would surface the implementation gap that written submissions cannot fully convey. The physician who carries the malpractice insurance and the regulator who decided no premarket standard was necessary should be in the same room answering the same questions. That conversation has not yet occurred at any level. The Dialogue can make it happen. These formats prioritize accountability over consensus and specificity over aspiration.

We offer four concrete examples, two already operational and two proposed in our written submission. Medical specialty board certification (operational for over a century). The most proven AI governance model predates AI entirely. Medical specialty boards administer validated, standardized competency assessments that a practitioner must pass before operating in a given clinical domain. This infrastructure - examination development, psychometric validation, pass/fail thresholds, periodic recertification - is directly applicable to Medical AI. Content-Driven Intelligence methodology has already demonstrated that AI systems can be tested against board-level standards using questions derived from the same source literature the AI accesses, cleanly separating retrieval capacity from reasoning capacity. The methodology exists. The mandate to apply it does not. Medical device classification frameworks (operational across jurisdictions). The FDA's 510(k), Premarket Approval, and EU MDR pathways require that medical instruments demonstrate safety and performance before reaching patients. These frameworks govern every device from surgical instruments to imaging systems. AI systems that inform individual clinical decisions function as medical instruments. Extending existing device classification to Medical AI requires regulatory will, not new infrastructure. The Lead Zero Standard (proposed, April 2026). Our companion governance protocol provides the operational specification for clinician authority in AI-assisted patient care. It replaces the human-in-the-loop model - where the AI acts first and the clinician reacts - with Clinician Starts the Loop, where every AI interaction begins with the clinician's declared intent. The standard has been stress-tested against 24 adversarial attack classes and scales from a rural clinic with a tablet to an advanced trauma center without modification to core principles. The AMA Governance Toolkit (operational, August 2025). The American Medical Association's eight-step framework provides institutional governance guidance for clinical AI deployment, representing the largest physician organization's formal recognition that this governance gap requires immediate action. Each example demonstrates the same principle: effective AI governance for high-consequence domains is built on existing infrastructure, not invented from first principles.