MICROSOFT

For a first global Dialogue on AI Governance to be successful it should show how to operationalize strategic thought, moving the conversation from principles to implementation. First, it should establish a shared baseline for AI risk. Today, there is no common understanding of systemic risks or how to manage them. Defining this baseline—including risks that emerge when AI interacts with other technologies—would be a meaningful step toward coherence. Second, it should clarify how coordination actually happens within the United Nations system. Not another layer of governance, but a practical mechanism—lightweight, evidence-driven—that helps align efforts, surface emerging risks, and avoid duplication. Third, the Dialogue should deliver operational guidance. The gap is no longer in principles, but in execution: how safety-by-design is implemented in real systems, how systemic risk assessments are conducted and validated, and how external expertise is integrated into product and policy decisions. Fourth, it should explicitly recognize public trust as a systemic variable. Algorithmic amplification can reinforce echo chambers, shaping perception and eroding trust. This directly affects adoption, contestation, and regulation of AI systems. Fifth, credibility will depend on who is meaningfully included. Multi-stakeholder participation from different organizations from academia, think tanks, and technology corporations across geographies to ensure inclusivity and a multi-faceted approach to help create an inclusive technology. Finally, success requires continuity. Not a one-off dialogue, but a process with clear follow-through, measurable outputs, and early signals of adoption across policy and practice.

• Safe, secure and trustworthy AI
• AI capacity-building
• Protection and promotion of human rights
• Interoperability of governance approaches

These priorities reflect where urgent action is needed to move from principles to implementation, and where our experience can contribute most directly. -Safe, secure and trustworthy AI is foundational. The key gap is not in defining principles, but in operationalizing them-through safety-by-design, real-time risk mitigation, and measurable safeguards in deployed systems. Advancing practical approaches to systemic risk management is critical. -Interoperability of governance approaches is equally urgent. Fragmentation across jurisdictions creates complexity for both regulators and industry. There is a clear need for aligned risk frameworks, shared definitions, and compatible assessment methodologies that allow different regimes to work together while respecting local priorities. -AI capacity-building is essential to ensure that governance is globally inclusive. Many regions lack the technical, institutional, and data capabilities required to engage effectively. Efforts should focus on practical enablement-tools, expertise, and knowledge transfer that support implementation, not just participation. -Protection and promotion of human rights underpins all of the above. As AI systems increasingly shape access to information and opportunities, ensuring fairness, accountability, and user protection is critical. This includes addressing systemic dynamics-such as how algorithmic amplification can reinforce information asymmetries-that directly impact societal outcomes. Taken together, these areas reflect a common objective: building governance frameworks that are actionable, globally coherent, and grounded in real-world deployment.

First, implementation and assurance. The gap is no longer in principles, but in how they are operationalized and validated. For example, there is limited convergence on how systemic risk assessments are conducted, audited, or benchmarked across organizations and jurisdictions. Second, public trust as a systemic variable. Trust is increasingly shaped by technology dynamics, where algorithmic amplification can reinforce echo chambers and influence how information is perceived. This directly affects how AI systems are adopted, contested, and regulated. Third, cross-domain risks. As AI combines with other technologies, capabilities compound and risk profiles shift. For instance, the intersection of AI with quantum computing or advanced cybersecurity capabilities could redefine both defensive and offensive risk landscapes, while governance frameworks remain largely siloed. Fourth, data and evaluation ecosystems. Effective governance depends on access to shared datasets and benchmarks. Today, there is limited coordination on common evaluation frameworks to test system performance, safety, and robustness across contexts. Finally, feedback loops between policy and deployment. AI systems evolve rapidly, but governance mechanisms to incorporate real-world evidence-such as incident reporting or external expert input-remain underdeveloped. Addressing these issues would strengthen the ability to translate existing priorities into practical, adaptive, and globally coherent governance frameworks.

These priorities reflect where urgent action is needed to move from principles to implementation, and where our experience can contribute most directly. Safe, secure and trustworthy AI is foundational. The key gap is not in defining principles, but in operationalizing them—through safety-by-design, real-time risk mitigation, and measurable safeguards in deployed systems. Advancing practical approaches to systemic risk management is critical. Interoperability of governance approaches is equally urgent. Fragmentation across jurisdictions creates complexity for both regulators and industry. There is a clear need for aligned risk frameworks, shared definitions, and compatible assessment methodologies that allow different regimes to work together while respecting local priorities. AI capacity-building is essential to ensure that governance is globally inclusive. Many regions lack the technical, institutional, and data capabilities required to engage effectively. Efforts should focus on practical enablement—tools, expertise, and knowledge transfer that support implementation, not just participation. Protection and promotion of human rights underpins all of the above. As AI systems increasingly shape access to information and opportunities, ensuring fairness, accountability, and user protection is critical. This includes addressing systemic dynamics—such as how algorithmic amplification can reinforce information asymmetries—that directly impact societal outcomes. Taken together, these areas reflect a common objective: building governance frameworks that are actionable, globally coherent, and grounded in real-world deployment.

The AI Dialogue can play a critical role by bridging the gap between global principles and coordinated implementation. First, it can serve as a neutral convening platform to align perspectives across governments, industry, and civil society. In a fragmented landscape, this is essential to move toward shared definitions of risk and compatible governance approaches, even where regulatory models differ. Second, it can advance practical interoperability. Rather than aiming for full harmonization, the Dialogue can help define common baselines—risk taxonomies, assessment methodologies, and safety expectations—that different jurisdictions can adopt and adapt. This would reduce friction while preserving regulatory sovereignty. Third, it can facilitate structured knowledge exchange grounded in real-world deployment. International cooperation is most effective when informed by operational experience—what works in practice, what fails, and how systems behave at scale. The Dialogue can institutionalize this exchange through case-based discussions and evidence-sharing. Fourth, it can strengthen inclusive participation by supporting capacity-building and ensuring that a broader set of countries can meaningfully engage in shaping AI governance, not only implementing it. Fifth, it can enable early coordination on emerging risks, particularly those that cross borders or sectors. By surfacing signals early and connecting stakeholders, it can help avoid reactive, uncoordinated responses. Finally, the Dialogue can create continuity and accountability, establishing a process with follow-up mechanisms and measurable outputs. Cooperation requires sustained engagement, not one-off alignment. In this way, the AI Dialogue can help shift international governance from fragmented and reactive efforts to a more coordinated, evidence-based, and implementation-driven approach.

On the normative side, UNESCO's Recommendation on AI Ethics and the work of the OECD provide widely recognized principles and policy frameworks. On the technical side, standards bodies such as ISO and IEEE are advancing methodologies for risk management, safety, and system evaluation. Multilateral and multi-stakeholder initiatives—such as the Global Partnership on AI and the International Telecommunication Union's AI for Good platform—contribute research, capacity-building, and convening power. In parallel, regional regulatory frameworks (e.g., the EU AI Act) are driving implementation at scale. The added value of the AI Dialogue lies in connecting these layers into a coherent, operational ecosystem. First, it can act as a coordination layer, linking principles, standards, and regulatory practices. Today, these efforts often evolve in parallel, with limited alignment on how they translate into implementation. Second, it can focus on interoperability in practice—aligning risk taxonomies, assessment approaches, and evaluation methods so that outputs from different initiatives can be compared and reused across jurisdictions. Third, it can provide a bridge between policy and deployment, bringing in real-world operational experience to inform governance. This includes structured input from practitioners on how safety measures, audits, and interventions perform in live systems. Fourth, it can strengthen continuity and feedback loops, ensuring that insights from implementation inform standards and policy evolution over time. In this way, the Dialogue's value is not to create new frameworks, but to make existing ones work together more effectively and translate them into practice.

Different stakeholders can contribute most effectively if the Dialogue is structured to capture both expertise and operational experience, not just positions. First, participation should be role-based and complementary. -Governments: articulate policy objectives and constraints -Industry: provide deployment experience, risk signals, and implementation insights -Academia and civil society: contribute independent analysis, evaluation methods, and societal impact perspectives -International organizations: ensure coordination and continuity Second, the Dialogue should combine three formats: -High-level plenaries to align on priorities and signal political commitment -Thematic working groups (e.g., safety, interoperability, capacity-building, human rights) focused on producing concrete outputs such as risk taxonomies, implementation guidance, or evaluation approaches -Practitioner roundtables with smaller, curated groups to discuss real-world cases, including system deployment, incident response, and mitigation strategies Third, it should include structured input mechanisms beyond invited participants. Open consultations, written submissions, and targeted expert interviews can broaden participation while maintaining quality. Fourth, the Dialogue should ensure meaningful inclusion of underrepresented regions, not only through participation but through capacity-building support that enables stakeholders to contribute substantively. Finally, the process should be iterative and outcome-driven. Each phase should produce clear deliverables, with feedback loops that incorporate lessons from implementation and evolving risks. In this structure, stakeholders are not only represented—they actively contribute to building practical, evidence-based governance approaches.

Several perspectives remain systematically underrepresented in global AI governance discussions, particularly those closest to real-world impacts and implementation. First, Global South policymakers and practitioners. While often represented at a high level, there is limited inclusion of technical and operational voices from these regions. This constrains the ability to reflect local constraints, priorities, and deployment realities. Inclusion requires targeted capacity-building, sustained funding, and mechanisms that support ongoing participation, not just one-off representation. Second, frontline practitioners—those designing, deploying, and monitoring AI systems (e.g., trust & safety, product, and risk teams). Their experience with real-world system behavior, adversarial dynamics, and mitigation effectiveness is often absent from policy discussions. Structured practitioner roundtables and case-based inputs can address this gap. Third, affected communities and end users, particularly vulnerable groups. Their perspectives are typically mediated through institutions rather than directly represented. More direct engagement—through participatory mechanisms, user research, and civil society partnerships—can provide grounded insights into impacts. Fourth, independent researchers and auditors with access to empirical evidence. There are still barriers to accessing data, tools, and systems needed to evaluate AI performance and risks. Expanding secure data-sharing frameworks and supporting independent evaluation ecosystems would strengthen their contribution. Finally, cross-domain experts (e.g., cybersecurity, biosecurity, emerging technologies such as quantum computing) are often not fully integrated, despite the growing importance of intersections between AI and other technologies. Including these voices requires moving beyond representation toward structured, sustained, and well-resourced participation, with clear pathways for their input to shape outcomes.

To foster meaningful and dynamic engagement, the Dialogue should move beyond traditional panels toward formats that surface real-world experience, test assumptions, and produce actionable outputs. First, case-based clinics. Small, curated sessions where practitioners present concrete scenarios (e.g., risk mitigation, incident response, system deployment) under structured discussion formats (e.g., Chatham House Rule). This enables candid exchange and produces practitioner-informed insights. Second, scenario-based exercises. Cross-stakeholder groups work through forward-looking situations—such as emerging risks or cross-border incidents—to stress-test governance approaches and identify coordination gaps. This is particularly useful for systemic and cross-domain risks. Third, co-creation sprints. Time-bound working sessions where participants jointly develop outputs—such as risk taxonomies, evaluation frameworks, or implementation guidance. This shifts engagement from discussion to tangible deliverables. Fourth, evidence and data exchanges. Structured sessions where participants share empirical findings—metrics, evaluations, or case studies—in a comparable format. This helps ground the Dialogue in real-world performance and outcomes. Fifth, practitioner–policy pairings. Deliberate pairing of policymakers with operators (e.g., product, safety, engineering) to bridge the gap between policy intent and system implementation. Finally, iterative engagement cycles. Rather than one-off sessions, the Dialogue should incorporate follow-ups where outputs are revisited, tested, and refined based on feedback and evolving evidence. These formats would help ensure the Dialogue is not only inclusive, but also practical, evidence-based, and outcome-driven.

Effective AI governance is increasingly driven by practices that connect policy with real-world system deployment. One example is the use of structured external engagement to inform product and policy decisions. Inviting experts from civil society and academia to present on trust & safety topics within product teams enables direct integration of independent perspectives into system design and risk mitigation strategies. A second approach is the use of expert clinics and practitioner roundtables, where external specialists and internal teams discuss concrete challenges-such as safety features, policy enforcement, or emerging risks. These formats support evidence-based decision-making and help validate approaches before and after deployment. Third, collaboration networks within the AI trust and safety community play a critical role. These networks enable practitioners across organizations to exchange insights, align on emerging risks, and develop shared approaches to complex challenges, particularly where issues are cross-platform. Fourth, data intelligence partnerships, including collaboration with external fact-checkers, help inform how systems assess and elevate authoritative content. Integrating these signals strengthens the ability to respond to misinformation and improve the reliability of outputs. Fifth, proactive engagement with regulators helps bridge the gap between policy intent and technical reality, contributing to more informed and implementable frameworks. Finally, governance is most effective when safety measures are embedded directly into products, supported by continuous feedback loops that combine internal metrics, external input, and real-world system behavior. Together, these approaches demonstrate that effective AI governance depends on the ability to operationalize, test, and continuously improve safeguards in practice.