Lisa Intelligence Systems Inc.

A successful outcome would be a clear shift from high-level AI governance principles toward practical, enforceable mechanisms. Today, many frameworks define what AI systems should do. However, fewer address how these expectations can be upheld once systems are deployed and operating autonomously in real-world environments. As AI systems evolve into agents capable of acting, interacting, and making decisions over time, governance must extend beyond development and evaluation into continuous oversight during operation. A key outcome would therefore be recognition that governance requires enforceability. This includes the ability to observe system behaviour in real time, apply constraints during execution, and intervene when necessary. Another important outcome would be stronger alignment between policy frameworks and technical implementation. Governance must be grounded in what can realistically be built and deployed across sectors. The Dialogue could also identify gaps between current governance approaches and emerging system capabilities, particularly in relation to autonomous and agent-based systems. Finally, success would include progress toward interoperable governance mechanisms that function across jurisdictions and technical environments.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Interoperability of governance approaches
• Protection and promotion of human rights

These priorities reflect the need to move from conceptual governance toward systems that can function reliably in practice. Safe, secure and trustworthy AI is essential, but trust ultimately depends on whether systems can be monitored and controlled once deployed. This connects directly to transparency, accountability, and human oversight, which require not only visibility into system behaviour but also mechanisms to act on it. Interoperability of governance approaches is increasingly important as AI systems operate across borders and technical environments. Without interoperability, governance risks becoming fragmented and difficult to enforce consistently. The protection and promotion of human rights remains central, especially as AI systems influence decisions at scale. Ensuring these rights are upheld requires governance mechanisms that are not only defined in principle, but operational in real-world conditions. Together, these priorities highlight the importance of bridging the gap between governance design and governance execution.

A key emerging issue is the need for governance that operates at runtime, not only at the stages of development and deployment. As AI systems become more autonomous and interconnected, risks increasingly arise during operation. These include behavioural drift, complex system interactions, and challenges in tracing and correcting decisions once processes are underway. Current governance frameworks do not fully address these dynamics, creating a gap between defined rules and actual system behaviour. Addressing this requires new forms of governance infrastructure that enable real-time observability, enforceable constraints, and timely intervention. Another important issue is the alignment between policy design and technical feasibility. Governance approaches must reflect how systems are actually built and operated, to ensure that requirements can be meaningfully implemented. These challenges suggest that the next phase of AI governance will depend not only on defining principles, but on ensuring they can be applied and maintained in practice.

The governance gaps in these areas are already affecting both sectors and jurisdictions by creating uncertainty around the safe deployment of increasingly capable AI systems. A major challenge is that AI governance remains stronger at the level of principles than at the level of implementation. Many organizations now understand the importance of safety, accountability, and human oversight, but they still lack practical mechanisms to ensure these requirements are upheld once systems are deployed. This is especially relevant in sectors such as finance, public services, and critical infrastructure, where AI systems may increasingly interact with sensitive data, operational processes, and high-impact decisions. Another challenge is fragmentation. Different governance approaches are emerging across jurisdictions, which creates uncertainty for organizations building or deploying AI across borders. Without interoperability, governance risks becoming inconsistent and difficult to apply in practice. At the same time, there is a major opportunity. The current moment allows governments, standards bodies, companies, and civil society to shape governance before more advanced autonomous systems become deeply embedded in institutions and markets. There is also an opportunity to move from abstract governance toward operational governance. If the international community can support mechanisms for observability, accountability, and enforceable constraints in real-world systems, this could strengthen trust, reduce risk, and make AI deployment safer and more sustainable across regions and sectors.

The AI Dialogue can play an important role by helping to bridge the current gap between fragmented national approaches and the growing need for globally interoperable AI governance. At present, many jurisdictions are developing their own frameworks, standards, and regulatory models. While this reflects legitimate differences in priorities and legal systems, it also creates a risk of fragmentation at a time when AI systems, infrastructure, and data flows increasingly operate across borders. The Dialogue can help establish a shared understanding of core governance challenges and identify areas where coordination is both possible and necessary. It can also serve as a forum for connecting policy discussions with technical realities. One of the key needs in AI governance is greater alignment between high-level principles and the mechanisms required to implement them in practice. By bringing together governments, technical experts, industry, academia, and civil society, the Dialogue can help ensure that international cooperation is grounded in both policy legitimacy and technical feasibility. In addition, the Dialogue can support confidence-building. Shared terminology, common reference points, and greater transparency around governance approaches can reduce mistrust and improve coordination across regions. Over time, the Dialogue could help create pathways toward interoperable standards, practical governance tools, and more consistent approaches to safety, accountability, and oversight. In that sense, it can become not only a venue for discussion, but a foundation for more durable international cooperation on AI governance.

The AI Dialogue should build on existing initiatives that have already created important foundations for international cooperation, while adding a more inclusive and implementation-oriented layer. Key reference points include the UN Global Digital Compact and the recommendations of the UN High-level Advisory Body on AI, which called for a global policy dialogue on AI governance. The Dialogue should also connect with the OECD AI Principles and the OECD/GPAI ecosystem, which provide widely used policy frameworks and analytical work on emerging AI risks. It should further build on the G7 Hiroshima AI Process, which has developed international guiding principles and a code of conduct for advanced AI systems, as well as the emerging network of AI Safety Institutes, which is strengthening international cooperation on AI safety. In addition, ITU's AI for Good process already brings together stakeholders across technical, policy, and standards communities. The added value of the AI Dialogue would be to connect these efforts in a more coherent way across the UN system and across stakeholder groups. It can provide a space where policy frameworks, technical standards, and operational realities are discussed together rather than in separate silos. In particular, the Dialogue could add value by helping translate existing principles into more interoperable and practical governance approaches, including for emerging challenges such as autonomous and agent-based systems. In that sense, it can serve as a bridge between existing frameworks and the next generation of implementable global AI governance.

Different stakeholders should contribute in ways that reflect their actual role in the AI ecosystem. Governments can bring regulatory experience, public-interest priorities, and international coordination capacity. Industry can contribute practical knowledge about deployment, technical constraints, and implementation challenges. Academia can provide independent research and long-term analysis. Civil society can help ensure that human rights, inclusion, and public accountability remain central. Standards bodies and technical experts can help translate governance goals into operationally meaningful mechanisms. To make this effective, the Dialogue should combine plenary discussions with smaller thematic working formats. High-level sessions are useful for political direction, but technical and operational issues require more focused exchanges. A good structure would therefore include: plenary sessions for common priorities, thematic roundtables, and multistakeholder working groups focused on concrete outputs. The Dialogue would also benefit from continuity between meetings. Rather than functioning only as a one-off event, it could support recurring tracks on key topics such as safety, interoperability, accountability, and governance of autonomous systems. Written submissions, targeted consultations, and follow-up working papers could help maintain progress between formal meetings. A successful structure should therefore combine inclusiveness with depth, and discussion with practical follow-through.

Several perspectives remain underrepresented in global AI governance discussions. One important gap is the relative absence of practitioners working on real-world deployment, including engineers, operators, safety architects, and those dealing with system behaviour in practice. Governance discussions are often rich in principles, but less informed by the realities of how AI systems are built, integrated, monitored, and controlled once deployed. Another underrepresented group includes actors from the Global South, especially those working in public institutions, local innovation ecosystems, and regional governance settings. Their perspectives are essential if global governance is to be legitimate and practically relevant across different contexts. Small and medium-sized enterprises, independent researchers, and public-interest technologists are also often less visible than large companies or well-resourced institutions. Yet they may bring important insights, especially on implementation barriers and emerging risks. In addition, communities directly affected by AI-enabled decisions, including workers, educators, public-service users, and marginalized groups, should be heard more consistently. Their experience can help ground governance discussions in actual social impact. These voices can be included through targeted outreach, funded participation support, hybrid formats, regional consultations, and structured speaking opportunities that do not depend only on institutional size or resources.

The Dialogue would benefit from formats that move beyond formal statements and allow participants to work through concrete governance problems together. One useful format would be scenario-based discussions. Participants could examine realistic cases involving advanced AI systems, such as autonomous agents in finance, public services, or critical infrastructure, and discuss what governance mechanisms would be required in practice. This would help connect abstract principles to operational realities. Another valuable format would be multistakeholder problem-solving sessions, where policymakers, engineers, civil society, and researchers work together on specific governance questions, such as accountability, intervention mechanisms, or interoperability across jurisdictions. Short technical-policy briefings could also help. These would allow participants to explain emerging issues in accessible terms and reduce the gap between policy language and technical implementation. Interactive written consultations before and after major meetings could further improve participation, especially for those unable to intervene live. This would also allow more thoughtful input from a wider range of stakeholders. Finally, the Dialogue could include recurring thematic tracks, rather than only general discussion, so that conversations can deepen over time and lead toward more practical outputs.

Effective AI governance is most credible when it combines policy principles, technical standards, and operational mechanisms. A useful example is the OECD AI Principles, which helped create a common international baseline around trustworthy AI. The EU AI Act is another important step because it translates broad governance concerns into a structured risk-based regulatory approach. The Hiroshima AI Process and the emerging network of AI Safety Institutes also provide valuable models for international coordination, safety testing, and shared technical learning. At the implementation level, good practice includes structured model evaluations, red-teaming, incident reporting, documentation, and clear human oversight requirements. These approaches are important because they help make governance more concrete and auditable. However, as AI systems become more autonomous, additional mechanisms are likely needed. Governance challenges increasingly emerge during operation, not only during development. This suggests that effective governance should also include technical capabilities such as real-time monitoring, traceability of system actions, enforceable policy constraints, and intervention mechanisms when systems behave unexpectedly. In other words, good governance may need to evolve from static compliance to operational governance. Another promising approach is interoperability between governance frameworks. Shared terminology, common reporting practices, and internationally understandable technical standards can reduce fragmentation and help make governance more consistent across jurisdictions. Taken together, the most effective approaches are those that connect high-level values with deployable technical and institutional mechanisms. That combination is essential if AI governance is to remain meaningful as systems become more capable, more autonomous, and more embedded in real-world environments.