Ariadne Labs |Brigham and Women's Hospital | Harvard T.H. Chan School of Public Health

AI governance can be a matter of life and death in health. The first Global Dialogue must deliver a concrete plan of action. It must produce an authoritative declaration that aligns AI policy with human rights and public health and sets specific targets, including auditable AI systems in critical health sectors by 2027. It must secure government commitments on common standards for health data and targeted financing for AI capacity-building in underserved regions. It must also establish an independent task force to oversee AI in healthcare. The outcomes must address urgent health risks directly. Most maternal deaths still occur in low and middle income countries, where access to timely, high quality care remains uneven. Regulators must validate AI diagnostic tools on local clinical data and audit their performance before scale. That is how governance prevents harm instead of formalizing it. WHO has also warned that climate change is already worsening risks for pregnant women and children. Health AI that ignores heat, pollution, and climate disruption will miss real drivers of risk. Inclusion and momentum matter. The outcome document must mandate multi-stakeholder working groups involving clinicians, patients, and technical experts to implement each commitment. It must set clear deadlines and connect to ongoing UN AI processes. This Dialogue must do more than state principles. It must lock in measurable action, named responsibility, and follow-through.

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

Trustworthy AI is non-negotiable. Unreliable systems directly harm patients. This demands rigorous safety standards and strong cybersecurity. Human rights must be protected at every stage. AI governance must uphold the rights to health, privacy, and non-discrimination, especially for women, children, and marginalized communities. Transparency and accountability matter just as much. Governments must require independent audits and clear reporting of AI decisions so clinicians and patients can judge whether these systems are safe, fair, and usable. The UN resolution establishing the Dialogue explicitly frames AI governance around international cooperation, human rights, and closing digital divides. Capacity-building is equally urgent. Many low-resource health systems still lack the infrastructure, workforce, and technical expertise needed to use AI safely and effectively. Without that foundation, innovation will deepen inequality instead of reducing it. AI fails fastest where systems are weakest. Global governance must start there. Together, these priorities make AI accountable to the people whose lives it will shape.

Climate resilience needs explicit attention. Extreme heat, air pollution, flooding, and other climate pressures already worsen maternal and child health, yet many AI governance frameworks still treat those risks as peripheral. Health AI models must incorporate environmental exposure data where it changes risk, triage, or access. Digital inclusion is another gap. AI tools must work for rural populations, people with disabilities, and linguistic minorities. A technically strong model still fails if people cannot be reached, cannot understand it, or cannot trust it. Global data governance also needs stronger treatment. Health AI depends on sensitive personal data. International standards for privacy, secure cross-border sharing, and public accountability are essential, especially during crises. Pandemic preparedness belongs here as well. COVID showed how quickly weak coordination turns into preventable harm. The Dialogue should establish clear expectations for AI-supported early warning, outbreak detection, and response planning in public health emergencies. Without these protections, AI governance becomes a technical exercise while people suffer the consequences.

In the health sector, especially across maternal, newborn, and child health, the biggest governance gap is not the absence of AI tools. It is the absence of rules strong enough to decide which tools are safe, clinically reliable, and fit for the settings where they are being deployed. WHO has warned that health AI can generate false, biased, or incomplete outputs and has called for stronger regulation, independent audits, and human rights safeguards. That gap is most dangerous in low-resource systems. In 2023, about 92% of maternal deaths occurred in low and lower-middle income countries. When AI enters systems already strained by workforce shortages, weak referrals, uneven data, and limited digital infrastructure, flawed tools can deepen the inequities they claim to solve. Climate stress makes this even more urgent. WHO has warned that heatwaves, floods, droughts, air pollution, and climate-linked infectious disease are already worsening pregnancy complications, preterm birth, low birthweight, and stillbirth. If AI governance ignores climate exposure, it will underestimate risk in exactly the populations already carrying the heaviest burden. The governance opportunity is real. Strong rules can force local validation, independent auditing, public accountability, and practical use in triage, decision support, surveillance, and early warning. They can also shift power. LMIC institutions must not remain testing grounds for imported systems. They must help shape the standards, data rules, and oversight mechanisms from the start. Global AI governance will matter only when the systems most at risk help write the rules.

The AI Dialogue can do what most existing processes have not done: force coherence in a field that remains fragmented, uneven, and too easy for powerful actors to shape in their own interests. The UN created it to support international cooperation and open discussion on AI governance. Its real value lies in drawing a harder line on what countries owe people when AI affects rights, safety, and access to essential services. Its most important role is practical. It must identify a small set of areas where international cooperation is not optional: cross-border data governance, independent auditing, safety standards for high-impact uses, public-interest compute and data infrastructure, and capacity-building for low and middle income countries. WHO's 2024 health AI guidance already points to regulatory oversight, post-release audits, and public infrastructure. The Dialogue must push that logic upward, so countries move toward comparable safeguards instead of fragmented and unequal ones. It must also correct a power imbalance. Too much AI governance still gets shaped by a narrow set of states and firms, while lower-resource countries are asked to absorb tools and rules they did not help design. The Dialogue can change that by bringing LMIC institutions, sector experts, civil society, and affected communities into agenda-setting from the start, not after standards are already written. Global AI governance will remain shallow until those most affected help write the rules.

The Dialogue should not start from scratch. It should build on work that already carries technical, ethical, and political weight. UNESCO's Recommendation on the Ethics of Artificial Intelligence offers a global foundation across human rights, accountability, transparency, environmental responsibility, and governance capacity. UNESCO's Readiness Assessment Methodology and Global Observatory already move that framework toward implementation. WHO's AI governance work for health is essential because it turns broad ethical principles into sector-specific expectations: regulatory approval pathways, mandatory audits, published impact assessments, and public-interest infrastructure. That matters in any discussion of high-impact AI. The OECD AI Principles, the OECD AI Policy Observatory, and the GPAI-OECD partnership add policy benchmarks, expert networks, and practical cooperation on trustworthy AI. ITU's AI for Good platform, along with standards work through ISO and IEC, matters where governance must translate into technical standards, interoperability, and operational use. The Global Digital Compact architecture, especially the Independent International Scientific Panel on AI, adds scientific grounding and UN system continuity. The Dialogue's added value is not duplication. It is bringing these strands into one accountable, globally visible process. It can connect ethical frameworks, technical standards, sector safeguards, and scientific advice under a forum with broader legitimacy and stronger pressure for follow-through. That is the missing layer.

The Dialogue must be built for decisions, not speeches. Plenary sessions can set direction, but the real work must happen in small cross-sector groups tasked with producing something concrete: draft standards, regulatory text, or time-bound action plans. This structure matters because health AI research already shows that trust breaks down when governance excludes patient and community input, lacks transparency, and weakens oversight. Each group must include governments, technical experts, sector practitioners, civil society, and people directly affected by AI systems. In health, clinicians, public health practitioners, and patient voices must sit alongside regulators and developers. Technical expertise matters, but it cannot dominate decisions that shape rights, safety, and access. The Dialogue also needs an implementation layer. Each thematic track must end with a published output, named responsibility, and a review point. Without monitoring and public reporting, it becomes another convening platform that produces language without consequence.

Those most affected by AI still have the least power in shaping its rules. The most underrepresented voices include researchers and public institutions from low and middle income countries, frontline health workers, women in caregiving roles, rural communities, linguistic minorities, people with disabilities, and communities already living with weak public systems. This is not only a representation gap. It is a governance failure. Evidence from health AI shows that systems trained or designed without diverse populations and local context can reproduce bias, deepen inequity, and fail when transferred across settings. Inclusion requires more than invitations. The Dialogue must give LMIC institutions formal roles in agenda-setting, drafting, and decision-making. It must fund participation, interpretation, accessibility support, and technical preparation. It must also run regional consultations before the main Dialogue so priorities are shaped locally rather than filtered through global centers of power. Affected communities must be represented inside each working track, not moved into symbolic side events. Inclusion must begin where priorities are set, language is negotiated, and standards are written.

The most effective engagement formats are the ones that expose governance failure in real terms and force participants to respond. One format is case-based review. Participants examine real examples where AI caused harm, failed to generalize across settings, or raised unresolved accountability problems. That approach is stronger than abstract panels because recent research shows recurring concerns around bias, black-box decision-making, weak oversight, and poor transferability across institutions. A second format is governance labs. Small mixed groups work through a live problem such as cross-border health data sharing, AI auditing in clinical care, or bias in public service delivery, and leave with draft text or a decision model. A third format is public commitment review. Governments, institutions, and companies state specific actions with timelines, publish them, and return to report progress. Public trust in health AI depends on visible oversight and accountability, not promises alone.

Effective AI governance already exists in parts. The strongest models do three things at once: they set clear rules, assign real oversight, and create tools that countries can actually use. WHO's recent guidance on AI for health is one of the clearest examples. It moves beyond abstract ethics and calls for regulatory approval pathways, independent post-release audits, published impact assessments, and public-interest data and computing infrastructure. That matters because high-impact AI needs scrutiny after deployment, not just before it. UNESCO's Recommendation on the Ethics of Artificial Intelligence remains one of the strongest global foundations. Its value is not only in the principles it sets across human rights, accountability, transparency, environmental responsibility, and governance capacity. It also gives countries practical tools to assess readiness and move from broad commitments to implementation. The EU AI Act offers a different strength: enforceable regulation. Its risk-based structure places stronger obligations on systems that affect health, safety, and fundamental rights, and it does not rely on voluntary compliance alone. The OECD AI ecosystem adds another layer through policy benchmarking, monitoring, and cross-country learning. That matters because AI governance fails when every jurisdiction works in isolation. The lesson is straightforward. Good AI governance does not come from principles alone. It comes from combining ethics, regulation, implementation tools, independent oversight, and public accountability. The AI Dialogue should build on that mix and press for alignment where fragmented rules now create the greatest harm.