Independent Researcher in AI Governance & Safety

A successful first Global Dialogue on AI Governance should do more than restate broad principles. It should identify a practical set of shared priorities that can guide implementation across different legal, institutional and technical contexts. First, it should establish a clearer common vocabulary for governance, especially around accountability, human oversight, risk management and institutional responsibility. Many current debates remain too abstract, which makes implementation uneven. Second, it should highlight a small number of actionable priorities for international cooperation, including capacity-building for regulators and public institutions, better interoperability across governance approaches, and stronger links between safety, accountability and operational oversight. Third, it should recognize that trust in AI cannot depend only on post-hoc evaluation of outcomes. A meaningful governance framework also requires attention to the conditions under which systems are deployed, authorized and monitored in practice. Finally, the Dialogue should produce a useful basis for follow-up: not only a summary of views, but a roadmap for continued exchange, practical learning and policy coordination among governments, international organizations, academia, technical communities and civil society. If the Dialogue can move from general agreement to clearer governance language, practical priorities and sustained cooperation, it will already be a meaningful success.

• Safe, secure and trustworthy AI
• AI capacity-building
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

I selected these four priorities because they are closely connected in practice. Safe, secure and trustworthy AI is an essential overall objective, but it becomes meaningful only when linked to transparency, accountability and human oversight. Without clear responsibility, review mechanisms and effective oversight, trust can easily remain rhetorical. Interoperability of governance approaches is also critical. AI systems, supply chains and policy effects cross borders, while governance frameworks remain fragmented. Greater interoperability is needed not to erase legal diversity, but to improve coordination, reduce gaps and support more consistent baseline protections. AI capacity-building is equally important because governance is implemented by institutions and people, not by principles alone. Many public authorities, regulators and oversight bodies still lack the technical, legal and operational capacity required to govern AI effectively. Taken together, these priorities reflect a practical view of AI governance: safety requires accountability, accountability requires institutional capacity, and all of them require better coordination across jurisdictions and governance models.

One important cross-cutting issue is the governance of authorization conditions before deployment or before a system is allowed to take consequential action. Many current governance discussions focus on outcomes: monitoring, auditing, transparency, redress and responsibility after an event has occurred. These are important, but they do not fully address the prior question of whether a system should have been permitted to operate, or to act, under those conditions in the first place. This is especially relevant for systems used in high-impact or governance-sensitive environments. In such contexts, meaningful oversight may require clearer ex-ante control points: defined authorization thresholds, deployment conditions, escalation rules, traceability requirements, and mechanisms for suspension or fail-safe intervention. This issue cuts across safety, accountability, human rights and interoperability. It is not only a technical question, but also an institutional one: who has the authority to permit, review, restrict or halt consequential AI use, and according to what standards. Greater attention to governance before consequential action would strengthen existing themes and make AI governance more operational.

In Japan and across the Asia-Pacific region, the main challenge is not only the rapid adoption of AI, but the uneven development of governance capacity around it. There is growing attention to safety, accountability and trustworthy AI, yet institutional readiness, operational oversight and cross-sector implementation remain inconsistent. One important gap is that governance often develops more slowly than deployment. Organizations may adopt AI systems before clear internal responsibility, review processes, escalation mechanisms or audit practices are in place. This creates practical uncertainty, especially in sectors where decisions can have legal, social or economic consequences. Another challenge is fragmentation. Different jurisdictions and sectors are moving at different speeds, with varying legal traditions, regulatory cultures and technical capacities. This makes interoperability difficult and increases the risk of uneven protections, weak accountability and confusion for both public institutions and private actors. At the same time, there is a major opportunity. Many countries in the region are still in a formative stage and can benefit from shared governance language, institutional learning and practical models for capacity-building. There is room to build governance approaches that are not only principled, but operational: with clearer responsibility, oversight, traceability and review mechanisms. From the perspective of an independent researcher in AI governance and safety, the most valuable opportunity is to strengthen the link between high-level principles and implementable institutional practice. That is where meaningful progress can still be made.

The AI Dialogue can play a useful role by serving as a practical coordination platform rather than only a high-level discussion forum. International cooperation on AI governance does not require full legal harmonization, but it does require greater clarity on shared problems, common governance language and areas where coordination is both possible and necessary. A valuable role for the Dialogue would be to help bridge different governance communities: governments, regulators, international organizations, academia, civil society and technical experts. In many cases, these groups identify similar risks but use different concepts, priorities and operational assumptions. The Dialogue can reduce this fragmentation. It can also support cooperation by identifying a limited set of governance functions that are widely relevant across jurisdictions, such as accountability, oversight, traceability, review mechanisms, institutional capacity and conditions for responsible deployment. These are areas where practical exchange may be more productive than abstract consensus. In addition, the Dialogue can help connect policy principles with implementation experience. International cooperation is often strongest when it focuses not only on what values should be protected, but also on how institutions can actually govern AI in practice. Its value, therefore, lies less in producing broad declarations and more in improving coordination, comparability and operational learning across diverse governance settings.

The AI Dialogue should build on existing international and regional work rather than duplicate it. Useful reference points include the UN system's broader digital cooperation efforts, OECD work on AI policy and governance, UNESCO's Recommendation on the Ethics of AI, standards development bodies, and regional regulatory initiatives such as those emerging in the European Union. It should also pay attention to practical governance efforts led by public institutions, research communities and technical standard-setting processes. Its added value would not come from replacing these initiatives, but from connecting them more effectively. Many existing mechanisms are strong in particular areas: principles, ethics, standards, regulation, technical guidance or policy analysis. However, they are often fragmented across institutions, sectors and regions. The AI Dialogue could provide value in three ways. First, it could function as a bridging space across governance traditions and institutional levels. Second, it could highlight practical lessons on implementation, including institutional capacity, oversight and accountability. Third, it could help identify common governance functions that can travel across jurisdictions even where legal systems differ. In that sense, the Dialogue's comparative advantage is not to create a single model of AI governance, but to improve coherence, coordination and practical learning across the models that already exist.

Different stakeholders contribute most effectively when they are not only invited to speak, but are given roles that match their practical knowledge. Governments and international organizations can share policy priorities and coordination needs; regulators and public institutions can contribute implementation lessons; academia can provide analytical frameworks and evidence; civil society can highlight rights and social impacts; and technical communities can clarify operational constraints, system behavior and feasible control points. To make this meaningful, the Dialogue should be structured around a combination of plenary discussion and smaller thematic working formats. Large sessions are useful for visibility and political signaling, but smaller moderated sessions are more effective for identifying concrete governance problems and practical lessons. The Dialogue would also benefit from a structure that distinguishes between principles, implementation and institutional capacity. These are often discussed together, but they raise different questions and require different expertise. A useful format would therefore include: 1.high-level sessions on shared priorities, 2.focused technical and institutional discussions on implementation challenges, 3.cross-stakeholder sessions on accountability, oversight and capacity-building, 4.short follow-up outputs that capture lessons, not only statements of position. Meaningful participation depends not only on inclusion, but on whether the structure allows different kinds of expertise to shape the discussion in a usable way.

Global discussions on AI governance often include governments, major companies, international organizations and established expert communities, but several important perspectives remain underrepresented. One underrepresented group is public-sector implementers: regulators, oversight bodies, procurement officials, inspectors, administrative decision-makers and others who must govern AI in practice. Their experience is essential because they face operational constraints that are often missed in high-level discussions. Another underrepresented perspective is that of actors working in governance-sensitive environments, including justice systems, education, health, social administration and local public services. These sectors often experience the governance consequences of AI most directly. Independent researchers and smaller institutional actors from outside major power centers are also frequently less visible, even when they offer valuable comparative or cross-cutting insights. In addition, voices from regions with more limited institutional capacity may be present formally but still have less influence on agenda-setting. Inclusion should therefore not rely only on open invitations. The Dialogue should actively create space for operational and regionally diverse perspectives through targeted outreach, balanced panel design, multilingual access, support for remote participation, and structured opportunities for written input to feed into the main discussions. Meaningful inclusion requires not only representation, but also formats that allow less prominent but practically important voices to shape outcomes.

The most effective formats are likely to be those that move beyond formal statements and allow participants to work through practical governance questions together. One promising format would be scenario-based workshops built around concrete governance problems, such as high-impact public-sector use, cross-border deployment, accountability failures or oversight challenges. These discussions often generate more useful insight than abstract exchanges. Another useful format would be short comparative sessions in which participants from different regions or sectors explain how they address a similar governance function, such as risk assessment, human oversight, traceability or institutional review. This could improve practical learning across different governance traditions. The Dialogue could also benefit from moderated "implementation clinics" where public institutions, technical experts and researchers discuss real institutional bottlenecks, including capacity gaps, unclear responsibility, lack of review mechanisms or weak coordination. To make engagement more dynamic, written submissions should not disappear into the background. They should be synthesized, grouped by theme and reflected back into the live discussions so participants can engage with emerging patterns rather than isolated comments. Innovative engagement is valuable when it produces usable learning. The goal should be not only to hear many voices, but to create structured interaction that improves governance understanding across communities and jurisdictions.

Useful examples of effective AI governance can already be found across different levels and traditions of governance. At the policy level, risk-based regulatory approaches, such as those emerging in the European Union, are important because they connect AI governance to concrete obligations, accountability and institutional oversight rather than relying only on general principles. OECD work has also been valuable in providing practical policy frameworks and shared vocabulary across jurisdictions. At the organizational level, concrete governance practices are equally important. These include impact assessment procedures, documented human oversight arrangements, traceability requirements, audit logging, review and escalation mechanisms, and clear allocation of responsibility across the lifecycle of AI use. In practice, governance becomes more effective when it is embedded into decision processes rather than treated as an external compliance exercise. At the technical and operational level, useful approaches include fail-safe design, staged deployment, access controls, post-deployment monitoring, and mechanisms that allow systems to be paused, reviewed or restricted when risk conditions change. Across these examples, the most valuable lesson is that effective AI governance depends on the connection between principles, institutions and operational control. Policies matter, but so do concrete practices, governance processes and technical design choices. Approaches that improve responsibility, traceability, reviewability and control before and during consequential use are especially promising.