QERA

A meaningful outcome would be a shift from principle-based governance to system-level design of responsibility. In real-world AI deployments, responsibility is no longer centralized — it is fragmented across layers, actors, and interactions. Governance frameworks must reflect this reality. This includes explicitly addressing multi-agent environments, and establishing mechanisms for evaluation, auditability, and accountability at the system level — not just at the level of models or organizations. Without this shift, governance risks remaining conceptually sound but operationally ineffective.

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches
• Protection and promotion of human rights
• Safe, secure and trustworthy AI

These priorities are critical because current AI governance approaches often remain fragmented and principle-based, while real-world AI systems are dynamic, multi-layered, and increasingly distributed. Ensuring safety, addressing societal impact, enabling interoperability, and protecting human rights all require a shift toward system-level governance that can operate across layers, actors, and interactions. Without this shift, governance risks remaining conceptually sound but operationally ineffective.

One cross-cutting issue that remains insufficiently addressed is the structural fragmentation of responsibility in AI systems. As AI systems evolve into multi-agent, multi-layered environments, decision-making is no longer attributable to a single model, organization, or actor. Instead, outcomes emerge from complex interactions across data, models, systems, infrastructures, and human inputs. In such settings, responsibility does not disappear - it becomes distributed, fragmented, and often opaque. This creates a fundamental gap between existing governance approaches and real-world system behavior. Many frameworks still assume relatively bounded systems and identifiable points of control, which no longer reflect how AI is deployed in practice. Addressing this issue requires a shift from actor-centric accountability to system-level design of responsibility. This includes developing mechanisms to trace, allocate, and maintain accountability across layers and interactions, as well as new approaches to evaluation and auditability that operate at the level of end-to-end systems. Without confronting this structural challenge, efforts across safety, human rights, and interoperability risk remaining conceptually aligned but operationally ineffective. This also raises deeper questions about how we define agency, control, and decision-making in increasingly autonomous systems.

Across regions and sectors, one of the most significant governance gaps is the mismatch between how AI systems are designed and how governance frameworks are structured. In practice, AI systems are increasingly multi-layered and involve multiple actors across jurisdictions — from data providers and model developers to system integrators and end users. However, governance approaches often remain fragmented, siloed, and based on assumptions of clear boundaries and centralized control. This gap creates several challenges. Responsibility becomes difficult to trace and enforce, particularly when outcomes emerge from interactions rather than isolated components. Regulatory approaches may also struggle to keep pace with rapidly evolving system configurations, leading to uncertainty for both developers and users. In some regions, this results in over-reliance on compliance-driven approaches, while in others, governance remains underdeveloped. At the same time, this gap creates an important opportunity: to move toward system-level governance frameworks that can operate across layers and actors. This includes designing clearer accountability structures, improving interoperability between governance approaches, and developing evaluation and audit mechanisms that reflect real-world system behavior. For sectors such as digital platforms, finance, and infrastructure, where AI is deeply embedded in decision-making, addressing these gaps is critical not only for risk management, but also for enabling sustainable and trusted innovation.

The AI Dialogue can play a critical role not only as a forum for exchange, but as a platform for aligning how AI governance is operationalized across different contexts. While there is growing convergence around high-level principles, significant gaps remain in how these principles are translated into practice — particularly in complex, multi-layered and multi-actor systems. The Dialogue can help bridge this gap by fostering shared understanding of how governance needs to function at the system level. In particular, it can support international cooperation in three ways. First, by enabling cross-regional alignment on governance approaches across layers — from data and models to systems and societal impact. Second, by advancing interoperability not only between regulatory frameworks, but also between different governance logics (technical, legal, organizational). Third, by creating space to address emerging challenges such as fragmented responsibility in multi-agent environments, where traditional accountability models no longer apply. Ultimately, the value of the Dialogue lies in its ability to move beyond coordination toward co-design — developing approaches that are not only globally relevant, but also practically implementable in real-world systems.

The AI Dialogue should build on and connect with a range of existing initiatives that have advanced AI governance from different perspectives. These include global normative frameworks such as the UNESCO Recommendation on the Ethics of AI, regulatory approaches such as the EU AI Act, and multi-stakeholder efforts like the OECD AI Principles and the Global Partnership on AI. In addition, standard-setting bodies such as ISO and IEEE play a critical role in operationalizing governance through technical standards. While these initiatives have made significant contributions, they often operate at different levels — normative, regulatory, technical — with limited integration across layers. This creates fragmentation in how governance is interpreted and applied in practice. The added value of the AI Dialogue lies in its ability to connect these efforts at the system level. Rather than duplicating existing work, it can serve as a platform to align governance logics across layers, enable interoperability between frameworks, and address emerging challenges such as distributed responsibility in multi-agent systems. By doing so, the Dialogue can help move from a landscape of well-developed but disconnected initiatives toward a more coherent and implementable global governance architecture.

Different stakeholders can contribute to the AI Dialogue not only by bringing diverse perspectives, but by engaging at different layers of the AI ecosystem. Policymakers, industry, technical experts, and civil society each hold partial views of how AI systems function and impact society. To be effective, the Dialogue should be structured in a way that connects these perspectives rather than treating them in isolation. This suggests a layered and interaction-based format. For example, discussions could be organized across key system layers — such as data, models, systems, and societal impact — while also enabling cross-layer exchanges where key governance challenges often emerge. In addition, the Dialogue should move beyond static panel discussions toward more interactive and design-oriented formats. This could include scenario-based sessions, cross-sector working groups, and iterative discussions focused on specific governance challenges, such as accountability in multi-actor systems or evaluation of end-to-end AI systems. Finally, continuity is critical. Rather than one-off events, the Dialogue should enable ongoing collaboration, with mechanisms to capture insights, test approaches, and refine governance frameworks over time. Such a structure would allow stakeholders not only to share perspectives, but to co-develop governance approaches that are both globally informed and practically implementable.

Global discussions on AI governance often underrepresent perspectives that are directly involved in the implementation and operation of AI systems, as well as those most affected by their outcomes. This includes system integrators, domain-specific practitioners (e.g., in healthcare, agriculture, and public services), and actors in the Global South whose contexts are not fully reflected in dominant governance frameworks. In addition, individuals and communities impacted by AI-driven decisions are frequently included only indirectly, if at all. A key reason for this underrepresentation is structural. Many global discussions are organized around institutional or disciplinary silos, which do not align with how AI systems function in practice — across layers, sectors, and actors. As a result, critical perspectives at the points of system integration and real-world use are often missing. To address this, inclusion should be designed into the structure of the Dialogue. This could involve creating dedicated roles for practitioners and affected communities within working groups, incorporating scenario-based discussions grounded in real-world use cases, and enabling participation beyond traditional policy and technical circles. More broadly, governance processes should move from representation as presence to representation as influence — ensuring that these perspectives meaningfully shape how governance frameworks are designed and implemented.

Traditional formats such as panel discussions and keynote sessions are valuable for sharing perspectives, but they are often insufficient for addressing the complexity of real-world AI governance challenges. In multi-layered and multi-actor AI systems, critical issues emerge not within isolated domains, but at the interfaces between technical, organizational, and societal components. This requires engagement formats that are interactive, cross-layer, and design-oriented. One promising approach is scenario-based engagement, where stakeholders collaboratively explore concrete use cases and governance challenges (e.g., accountability in multi-actor systems or evaluation of end-to-end AI decision processes). This helps ground abstract discussions in operational realities. Another approach is cross-layer working sessions, structured around key system layers (such as data, models, systems, and societal impact), combined with facilitated exchanges across these layers. This can surface gaps and misalignments that are otherwise difficult to identify. In addition, iterative formats — such as ongoing working groups or "design sprints" — can support the co-development and refinement of governance approaches over time, rather than relying on one-off discussions. Ultimately, the most effective formats are those that move from discussion to co-design, enabling stakeholders not only to exchange views, but to jointly develop governance approaches that are both globally relevant and practically implementable.

Several existing policies and practices provide valuable building blocks for more effective AI governance, particularly where they move beyond principles toward operationalization. For example, the EU AI Act introduces a risk-based regulatory framework that links obligations to specific use cases, helping translate governance into actionable requirements. Similarly, the OECD AI Principles have played an important role in establishing shared normative foundations across countries. At a more operational level, standards developed by ISO and IEEE contribute to embedding governance into technical processes, such as risk management, system quality, and lifecycle considerations. In addition, emerging practices such as AI audits, impact assessments, and model evaluation frameworks are helping to bridge the gap between policy and implementation, particularly in sectors like finance and digital platforms. What these approaches highlight is that effective AI governance requires alignment across multiple layers - from high-level principles and regulation to technical standards and operational practices. The key challenge, and opportunity, is to better connect these elements into coherent, system-level governance frameworks that can function across actors, contexts, and increasingly complex AI systems.