CROSS Global Research & Strategy

A successful first Global Dialogue on AI Governance should yield practical alignment suitable for real-world deployment, especially in high-stakes settings such as healthcare. I offer this perspective as a physician executive working at the intersection of clinical care, regulatory policy, and AI governance. In healthcare, governance is not abstract. It determines whether AI systems can be deployed safely, equitably, and credibly in settings where decisions affect diagnosis, treatment, access, workflow, trust, and outcomes. In my view, success would include three outcomes. First, the Dialogue should establish shared baseline concepts for safe, secure, and trustworthy AI, including transparency, accountability, human oversight, risk evaluation, and interoperability. Without a common vocabulary, governance frameworks will remain difficult to compare or coordinate. Second, the Dialogue should identify practical pathways for interoperability across legal systems, sectors, and regions. The goal should not be uniformity. The goal should be interoperability with legitimacy: enough shared structure to support cross-border cooperation, and enough local adaptability to reflect different cultures, capacities, and institutional realities. Third, the Dialogue should produce tools that Member States and stakeholders can use. One example is a Context and Transferability Assessment for AI Systems. Such a tool could help determine whether an AI system developed or tested in one setting can be responsibly used in another. This is critical in healthcare, where performance and safety may vary across populations, languages, data environments, infrastructure, and oversight capacity. The first Dialogue will succeed if it leaves participants with concrete mechanisms for comparing governance approaches, identifying gaps, and supporting safe deployment across different contexts. It should help AI governance become more coherent, more usable, and better grounded in the realities of the people and systems it is meant to serve.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

I selected these priorities because they are central to whether AI governance can work in high-stakes, real-world settings such as healthcare. Safe, secure, and trustworthy AI cannot be achieved through technical performance alone. In healthcare and other high-stakes domains, AI systems must be evaluated for their impact on safety, access, workflow, trust, equity, and outcomes. The social, economic, ethical, cultural, linguistic, and technical implications of AI are urgent because AI systems do not enter neutral environments. A system developed in one country, language, data environment, or health system may not perform safely or fairly in another. Governance must therefore account for context, transferability, and the conditions under which deployment is appropriate. Interoperability of governance approaches is also essential. AI systems often move across borders more easily than governance frameworks do. The goal should not be a single rigid global model, but compatibility among approaches so that countries and institutions can compare risks, share lessons, and coordinate oversight. Finally, transparency, accountability, and human oversight are necessary to sustain public trust and protect people in settings where AI may influence consequential decisions. Stakeholders should know how systems are evaluated, who is responsible for harms, when human review is required, and how systems are monitored after deployment. Together, these priorities support a practical governance agenda: one that is technically sound, legally grounded, context-aware, and usable by institutions making real deployment decisions.

Yes. One cross-cutting issue that deserves clearer attention is context and transferability. Many AI systems are developed, trained, or validated in one setting and then deployed in another. The listed themes address safety, trustworthiness, human rights, capacity-building, and interoperability, but they do not fully capture the practical question of whether an AI system can responsibly "travel" across populations, languages, institutions, legal environments, and resource settings. This issue cuts across both opportunity and risk. A system may appear safe and effective in one data environment, but perform differently when used with another population, workflow, infrastructure, or oversight model. This is especially relevant in high-stakes domains where AI may influence access to services, allocation of resources, or decisions affecting health, rights, and opportunity. The Global Dialogue could address this by encouraging a Context and Transferability Assessment for AI Systems. Such an assessment would ask where a system was trained, where it was tested, for whom it was designed, what populations or languages may be underrepresented, what local infrastructure is required, and what oversight mechanisms must be in place before deployment. A second emerging issue is post-deployment learning. Governance often focuses on pre-deployment review, but AI systems can change over time as data, users, environments, and incentives shift. The Dialogue should consider how institutions monitor performance, report harms, update systems, and withdraw systems when needed. Together, context, transferability, and post-deployment learning would help ensure that AI governance is not limited to principles or initial approvals, but remains responsive to real-world use across diverse settings.

In the United States, one of the most significant challenges is fragmentation. AI governance is developing across federal agencies, state legislatures, standards bodies, accreditors, health systems, professional societies, courts, and private companies. This activity is necessary, but it also creates uneven expectations for organizations trying to govern AI responsibly. In healthcare, I have seen this firsthand through my work on the team that co-developed URAC's healthcare AI accreditation standards. Standards are an essential layer of governance because they help define expectations for responsible AI use. But standards alone are not enough. The harder task is operationalizing them across institutions with different resources, workflows, patient populations, technology stacks, and governance maturity. This is where many U.S. healthcare organizations face a practical gap. They may recognize the need for AI governance, but still struggle with questions such as: Who owns oversight? What evidence is required before deployment? How should tools be monitored after implementation? When should patients or clinicians be informed? How do organizations align internal governance with state laws, federal policy developments, accreditation expectations, and evolving technical standards? The opportunity is to move from isolated policies to integrated governance systems. Healthcare organizations need governance models that connect standards, regulatory requirements, procurement, clinical validation, data governance, risk management, human oversight, and post-deployment monitoring. The United States has strong assets: research capacity, health system innovation, technical expertise, accreditation infrastructure, and active policy development. The challenge is to connect these assets into governance approaches that are consistent enough to build trust but flexible enough to work across different care settings. If done well, the U.S. healthcare sector can help demonstrate how AI governance can move from written standards to operational accountability in high-stakes environments.

The AI Dialogue can advance international cooperation by serving as a practical coordination mechanism for governance efforts currently developing in parallel. Its role should not be to replace national, regional, or sector-specific approaches. Different jurisdictions will need governance frameworks that reflect their laws, institutions, cultures, and capacities. The Dialogue's value lies in helping these approaches become more compatible, comparable, and usable across borders. It can do this in three ways. First, the Dialogue can help establish shared terminology and baseline expectations for safe, secure, and trustworthy AI. This would make it easier for countries and institutions to compare approaches, identify gaps, and avoid unnecessary duplication. Second, it can create a forum for case-based learning. Rather than focusing solely on broad principles, the Dialogue should examine real-world deployment scenarios, including instances where AI systems have succeeded, failed, or caused unintended harms. This would allow Member States and stakeholders to learn from implementation, not only policy design. Third, the Dialogue can support the development of practical tools for governance interoperability. Examples could include an interoperability map of existing frameworks, a context and transferability assessment, and guidance on post-deployment monitoring and accountability. International cooperation will be strongest if the Dialogue connects technical standards, legal frameworks, human rights obligations, and implementation realities. It should also ensure that countries with fewer resources are not only recipients of AI systems or governance models but also active contributors to their development. The Dialogue can help build a governance environment where AI innovation can travel across borders more responsibly, with clearer expectations for safety, accountability, and local adaptation.

The AI Dialogue should build on existing initiatives while adding a global coordination layer that those initiatives cannot provide on their own. From a U.S. healthcare perspective, relevant efforts include the Coalition for Health AI, which is developing consensus-based guidance for responsible health AI adoption across health systems, clinicians, regulators, payors, and industry; Stanford's Institute for Human-Centered AI, which contributes policy research, convening, and global AI governance analysis; and NIST's AI Risk Management Framework, which provides a practical structure for identifying and managing AI risks. In healthcare specifically, the Office of the National Coordinator for Health IT's HTI-1 rule has also raised transparency expectations for predictive decision-support tools in certified health IT. The Dialogue should also connect with standards-setting, accreditation, and assurance initiatives, including those emerging from healthcare accreditors such as URAC, technical standards bodies, professional societies, civil society, and cross-disciplinary experts. Physicians, nurses, patients, operational leaders, ethicists, data scientists, and non-technical institutional leaders should be part of this work because AI governance cannot be reduced to technical performance alone. The added value of the AI Dialogue is not to duplicate these efforts, but to connect them. It can compare approaches across regions, identify where governance frameworks already align, and highlight where gaps remain. It can also help translate sector-specific lessons, such as those from healthcare, into broader guidance for other high-stakes domains. Most importantly, the Dialogue can create a neutral forum for interoperability: linking technical standards, legal obligations, human rights principles, accreditation pathways, and implementation experience. This would help countries and institutions move from isolated initiatives toward governance approaches that are compatible across borders, adaptable to local contexts, and usable in real deployment decisions.

Different stakeholders should contribute through a structure that is accessible, continuous, and designed for participation across resource settings. The AI Dialogue should include Member States, civil society, technical experts, clinicians and other sector experts, academia, industry, standards bodies, youth representatives, and affected communities. Participation should not be limited to those with the resources to travel, speak English fluently, or engage through highly technical policy forums. I recommend a hybrid structure with several components. First, the Dialogue should offer virtual participation across time zones, with interpretation, captioning, accessible materials, and multiple ways to submit input. Written comments, short video statements, surveys, and regional listening sessions can help include participants who may otherwise be digitally or geographically excluded. Second, in-person convenings should rotate across regions and include travel or participation support where feasible, especially for civil society, patient or community representatives, youth participants, and experts from low- and middle-income countries. Third, the Dialogue should maintain a continuous feedback cycle, not only through annual meetings. This could include quarterly virtual sessions, thematic working groups, and structured consultations on draft outputs. Fourth, there should be an annual global convening to report findings, compare regional experiences, and allow stakeholders to present what is working, what is failing, and what support is needed. Finally, the Dialogue should produce public-facing materials, including white papers, implementation briefs, short-form educational videos, infographics, and multilingual summaries. These outputs should be written for both policymakers and non-technical audiences. The most useful structure would combine formal diplomacy with practical learning. Case-based sessions, regional roundtables, and cross-sector working groups would allow the Dialogue to capture lived experience, technical expertise, and governance lessons from different parts of the world. This would make participation broader, more equitable, and more useful for policy implementation.

Several voices remain underrepresented in global AI governance discussions, especially those closest to real-world deployment and those most likely to experience harm. These include patients, caregivers, frontline health workers, community-based organizations, disability advocates, youth, older adults, indigenous communities, linguistic minorities, migrant communities, and people living in low-resource or digitally excluded settings. In healthcare, nurses, primary care clinicians, social workers, patient navigators, and operational leaders are often closer to implementation than senior policy or technical teams, yet their perspectives are not consistently included. Small and mid-sized enterprises, public-interest researchers, local innovators, and institutions from low- and middle-income countries are also underrepresented. Many lack the resources to attend global convenings or participate in technical standards processes, even though they are directly affected by governance decisions. These groups can be included through deliberate design. The AI Dialogue should offer multilingual participation, interpretation, captioning, accessible materials, and multiple engagement formats, including written comments, short video submissions, surveys, regional listening sessions, and community consultations. Participation should not depend on the ability to travel, navigate complex policy language, or contribute through highly technical forums. The Dialogue should also support participation from under-resourced groups through travel support, stipends, regional hubs, and partnerships with trusted local organizations. In high-stakes sectors such as healthcare, affected communities should be included before systems are deployed, not only after concerns arise. Finally, the Dialogue should create roles for non-technical contributors. AI governance requires technical expertise, but it also requires clinical judgment, lived experience, cultural knowledge, legal insight, and operational understanding. Inclusion should be measured not only by who is invited, but by whether their input shapes the outputs.

Innovative engagement should move beyond traditional panels and create formats that allow participants to test governance ideas against real deployment challenges. One effective format would be case-based governance labs. Participants could examine realistic AI deployment scenarios, such as an AI triage tool in a hospital, a public benefits eligibility tool, or a multilingual education platform. Each case could be reviewed from legal, technical, ethical, operational, and community perspectives to identify where governance succeeds or fails. A second format would be regional implementation roundtables. These sessions would allow countries and communities to share what is working in their contexts, the barriers they face, and the support they need. This would help prevent the Dialogue from being shaped only by high-resource settings. A third format would be interoperability workshops, where participants compare existing governance frameworks and identify areas of alignment, divergence, and potential compatibility. This would make the discussion more concrete and useful for policymakers. Finally, the Dialogue could support rapid synthesis briefs after each session, summarizing key lessons, unresolved questions, and proposed actions. These briefs should be public, accessible, and translated where feasible. The most effective engagement formats will be those that combine diplomacy with implementation: not only asking what principles should guide AI governance, but testing how those principles work across different sectors, populations, legal systems, and resource settings.

A useful set of examples comes from practices that make governance operational inside institutions, not only from broad policy frameworks. One example is accreditation-based governance. In healthcare, accreditation pathways can help translate responsible AI principles into reviewable organizational practices, including leadership accountability, risk management, data governance, validation, monitoring, and patient safety processes. This is valuable because it gives institutions a concrete way to demonstrate that AI oversight is not ad hoc. A second example is an AI system inventory or registry. Organizations should know which AI tools they use, where they are deployed, who owns them, what data they rely on, what risks they pose, and how they are monitored. Without an inventory, accountability is difficult. A third practice is pre-deployment review combined with post-deployment monitoring. AI governance should not end when a tool is approved. Institutions need processes to monitor performance drift, bias, safety events, user concerns, and workflow changes over time. A fourth approach is the assessment of context and transferability. Before adopting an AI system developed elsewhere, organizations should evaluate whether it fits their population, language, infrastructure, legal obligations, and oversight capacity. Fifth, procurement requirements can be powerful. Buyers should require vendors to provide documentation on training data, validation, intended use, limitations, human oversight, cybersecurity, incident reporting, and update practices. Finally, case-based governance reviews can help institutions learn from real deployment scenarios. These reviews can bring together technical experts, clinicians, legal teams, operational leaders, and affected communities to examine how AI performs in practice. Together, these approaches show that effective AI governance requires more than principles. It requires institutional routines, documentation, accountability, monitoring, and feedback loops that can be adapted across sectors and settings.