Causal Foundry

A successful first Global Dialogue would deliver three concrete outcomes. First, a practical shared agenda for international cooperation focused on a limited number of areas where coordination is both urgent and feasible — moving from high-level principles to implementable guidance on safe, transparent, and equitable AI deployment, particularly in high-stakes sectors like health. Second, genuine capacity-building pathways ensuring that low- and middle-income countries help shape governance rather than inherit it. This means strengthening technical expertise, institutional readiness, and access to governance tools that support responsible local deployment and meaningful oversight. Third, an inclusive, evidence-based process that centers operational experience — not only from regulators and large technology firms, but from organizations deploying AI in real public-service environments. Frontline implementers hold critical knowledge about maintaining human oversight, managing risk in low-resource settings, and adapting systems to local languages and infrastructure realities. Success means the Dialogue produces a credible roadmap for action, establishes mechanisms for ongoing collaboration, and ensures that countries and communities currently underrepresented in AI governance can meaningfully shape the next phase of global rule-setting — not merely ratify decisions made elsewhere.

• Safe, secure and trustworthy AI
• AI capacity-building
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

These four themes directly reflect the work of organizations building and deploying AI in high-impact, real-world settings. Safe, secure and trustworthy AI is foundational. Systems used in health and public-service delivery must be reliable and robust under real-world constraints - limited infrastructure, fragmented data, and users who depend on frontline workers rather than specialists. Safety cannot be designed only for high-resource environments. AI capacity-building is essential because meaningful adoption requires more than access to tools. Countries and institutions need technical expertise, practical governance workflows, and local ownership of implementation - not just training materials or borrowed frameworks. Protection and promotion of human rights matters because poorly designed or contextually misaligned AI can deepen exclusion. In digital health especially, marginalized populations face existing barriers to care; AI that is not designed for equity, accessibility, and local relevance risks compounding those inequities. Transparency, accountability, and human oversight are non-negotiable in any system informing care pathways, resource prioritization, or public-service decisions. AI should augment human judgment, not replace it, and its use must remain auditable and intelligible to the institutions responsible for deployment.

Three issues deserve more explicit attention. The first is contextual fit and operational realism. Governance discussions often center on frontier models or national regulation, while most implementation challenges arise at the deployment layer: intermittent connectivity, limited devices, multilingual environments, fragmented data systems, and variable institutional capacity. Frameworks that ignore these realities will fail the communities that need governance most. The second is evidence generation for deployed systems. Beyond principles, there is an urgent need for practical norms governing how to evaluate real-world impact, monitor unintended effects, and update systems responsibly over time - particularly for adaptive systems that evolve based on user behavior or new data. One-time approval processes are insufficient. The third is equity in access to infrastructure, data, and implementation pathways. Even when models are openly available, unequal access to compute, high-quality local data, and technical talent can reproduce structural imbalances. A meaningful global dialogue must ask not only who sets the rules, but who can realistically build, adapt, audit, and benefit from AI systems. These issues cut across safety, rights, transparency, and capacity-building, and are especially consequential in sectors like health where implementation quality directly shapes human outcomes.

In digital health, governance gaps create both friction and missed opportunity. The most pressing challenges are uneven institutional capacity, fragmented standards, and the absence of practical guidance on governing AI systems that support real-world decisions. In many low- and middle-income settings, organizations face genuine demand for AI-enabled tools but lack clear frameworks for risk documentation, human oversight, procurement standards, or evaluation of adaptive systems. This slows responsible adoption, creates hesitation among public-sector partners, and undermines the ability to scale solutions across countries. Equity gaps compound these challenges. Governance approaches designed primarily around high-resource contexts often overlook multilingual needs, data scarcity, infrastructure constraints, and the critical role of community health workers. In health, these omissions matter: weak governance can reinforce exclusion and erode trust in digital systems at the moment when that trust is most needed. The opportunities, however, are substantial. Better international coordination could establish practical deployment standards, improve interoperability, and build local capacity to govern and use AI effectively. In our sector, this would accelerate responsible use of AI for clinical decision support, patient engagement, workforce assistance, and supply-chain efficiency — and help ensure that institutions retain accountability and human control over consequential decisions.

The AI Dialogue can serve as a critical bridge between principles, policy, and implementation. First, it can create a trusted multilateral space where governments, technical experts, civil society, and implementers identify areas where common approaches are both necessary and achievable. In domains such as safety, transparency, evaluation, and human oversight, fragmented governance increases costs and uncertainty for everyone — particularly for smaller institutions in lower-resource settings. Second, it can translate broad commitments into practical cooperation: shared terminology, voluntary guidance, implementation toolkits, peer learning mechanisms, and accessible technical assistance. For many countries and organizations, this kind of operational support matters as much as high-level norm-setting. Third, the Dialogue can ensure that international cooperation reflects deployment realities in public-interest sectors — health, education, social protection — by including organizations that build and use AI in real service environments, not only those shaping frontier model policy. Finally, it can counter the risk of a fragmented global landscape where standards are set by a small number of actors and adopted by others without meaningful participation. The Dialogue's distinctive value is its potential to create a more inclusive, implementation-aware, and globally legitimate foundation for AI governance cooperation.

The Dialogue should build on existing multilateral and sectoral efforts rather than duplicate them — including the Global Digital Compact, mechanisms under resolution 79/325, and ongoing work on digital public infrastructure, human rights, and responsible AI in health, education, and humanitarian contexts. It should also connect with implementation-focused partnerships already operating across governments, civil society, researchers, and technology providers in low- and middle-income countries. These collaborations generate the most actionable lessons: how AI functions in practice, how to integrate it into workflows, how to manage risk in under-resourced environments, and what genuine human oversight looks like at scale. The Dialogue's added value is synthesis and elevation. It can connect fragmented efforts into a more coherent global process, surface practical lessons from diverse implementation contexts, and align normative discussions with operational realities. In doing so, it can become the forum that credibly links principles, technical cooperation, and real-world learning — giving the global governance process legitimacy that no single initiative currently commands.

Different stakeholders should contribute in complementary ways. Governments should articulate policy priorities, institutional constraints, and cooperation needs. Technical experts should provide evidence, evaluation methods, and practical guidance on safety and oversight. Civil society should surface rights risks, inclusion concerns, and affected community perspectives. Implementation partners should contribute operational lessons — including failures — that reveal what governance requires in practice. The Dialogue's structure should combine a high-level political track with a more technical and implementation-oriented one. Plenary sessions can set direction, but thematic working sessions are needed to produce substantive outputs. Crucially, the structure must accommodate contributions from organizations outside major policy centers, including those deploying AI in low-resource settings. A productive format would include multistakeholder plenaries, focused thematic roundtables, structured written submission processes, and clear mechanisms for distilling recommendations, open questions, and options for future cooperation. The process should be iterative — not a single event — so that lessons from practice continuously inform governance discussions and emerging issues can be revisited as they develop.

Several perspectives remain systematically underrepresented: implementers from low- and middle-income countries, frontline public-service workers, smaller technology developers working in the public interest, local researchers, and communities directly affected by AI-enabled decisions. In health, community health workers, local administrators, and organizations deploying AI in everyday service settings hold irreplaceable knowledge about real governance questions — when systems fail, how oversight functions in practice, what users trust, and what barriers exist to safe deployment. Yet these voices are rarely as visible as major governments, global firms, or large research institutions. Inclusion requires structural accommodation, not just broader invitations. This means regional consultations, accessible remote participation, multilingual materials, travel support where needed, and structured opportunities for written practitioner input. The Dialogue should also create dedicated space for sector-specific implementation experience — from health, education, and humanitarian settings — rather than treating governance as primarily a legal or geopolitical matter. A process that aims to produce globally relevant outcomes must ensure that those living with the consequences of governance choices can meaningfully shape them.

The most effective formats would combine breadth of participation with concrete problem-solving. Scenario-based policy workshops would bring together mixed groups — policymakers, technologists, civil society, and sector practitioners — to work through real governance cases: AI use in health triage, community health worker support, or public procurement. Grounding discussions in specific trade-offs produces more actionable insight than debating abstract principles. Implementation showcases would give organizations deploying AI in different regional and resource contexts the opportunity to share operational lessons on oversight, risk mitigation, multilingual adaptation, and low-connectivity environments. This would make the Dialogue genuinely useful to countries building governance capacity. Regional listening sessions feeding structured input into the global meeting would ensure the Dialogue reflects distributed experience rather than only what surfaces in a single room. Synthesized outputs from these sessions should visibly shape plenary discussions. Cross-sector challenge sessions could focus participants on a small number of shared governance problems — documentation standards, community inclusion, public-sector procurement guidance — where cross-domain collaboration produces practical solutions. Together, these formats would make the Dialogue participatory, evidence-grounded, and oriented toward outcomes.

Effective AI governance is built through operational practices that combine safeguards, documentation, and continuous learning. Human-in-the-loop deployment ensures AI supports decisions without replacing accountable human judgment - essential in health and public-service settings where professionals must remain responsible for consequential actions. Evidence-based evaluation of deployed systems goes beyond one-time approval. It includes continuous monitoring of outcomes, subgroup effects, and unintended consequences - particularly important for adaptive systems that evolve based on new data or user behavior. Governance frameworks should require mechanisms to test, document, and review system changes over time. Interoperable, auditable system design enables institutions to govern AI use responsibly. Platforms that are modular, versioned, and transparent about models, datasets, and decision logic lower the barrier to meaningful oversight. This includes controlled data flows, policy-driven decisioning, and auditable workflows as operational building blocks. Capacity-building embedded in implementation ensures institutions develop not just the ability to use AI tools, but to understand, monitor, and adapt them. Governance is most durable when local institutions build genuine ownership over both deployment and oversight - not dependency on external providers. Together, these practices demonstrate that responsible AI is not only a policy aspiration but an achievable operational standard.