Independent Contributor

A successful outcome of the first Global Dialogue on AI Governance would be the transition from high-level principles to operationally actionable foundations. While there is already broad consensus on the importance of safe, secure, and trustworthy AI, the key gap lies in how these principles can be consistently implemented and interpreted across different organizational and national contexts. Three outcomes would indicate meaningful progress: First, the emergence of a shared structural language for describing AI systems. This includes standardized ways to represent use cases, responsibilities, risk levels, and system behaviors. Without such a foundation, governance remains fragmented and difficult to align across domains. Second, agreement on core building blocks for operational governance. This may include elements such as AI asset identification, traceability mechanisms, and governance checkpoints embedded within system lifecycles. These components enable governance to function as part of execution, rather than as an external review process. Third, initial alignment on measurable indicators of governance effectiveness. To move beyond compliance, governance systems must be evaluated based on their ability to ensure reliability, accountability, and real-world impact. Even partial consensus on metrics would significantly advance comparability and interoperability. Ultimately, success would not be defined by the number of principles agreed upon, but by the establishment of practical, structurally compatible approaches that can be adopted, tested, and iterated across diverse contexts.

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

My selected priorities reflect a practical focus on making AI governance operational, scalable, and interoperable in real-world environments. Safe, secure and trustworthy AI represents the overall objective. However, in practice, trust cannot be achieved through principles alone-it must be grounded in how AI systems are structured, monitored, and controlled throughout their lifecycle. Transparency, accountability, and human oversight are essential to operationalizing this goal. In enterprise settings, this often requires establishing clear system descriptions, defined ownership, and embedded control points to ensure that AI systems remain understandable, auditable, and subject to appropriate human intervention. Interoperability of governance approaches is a critical but underdeveloped area. While many organizations have implemented internal governance mechanisms, these are often not structurally aligned, making it difficult to compare, integrate, or scale governance across domains. Advancing interoperability requires shared ways of representing AI systems and governance-relevant information. Finally, AI capacity-building is necessary to bridge the gap between governance design and implementation. Governance systems are only effective if organizations across different functions can consistently apply them in practice. Together, these priorities emphasize a shift from high-level alignment toward practical governance systems that are structurally consistent, operationally embedded, and adaptable across contexts.

The Missing Link: Lifecycle "Sunset" Policies and Dynamic Asset Re-evaluation While the current clusters cover the "Development-to-Deployment" phase extensively, they largely overlook the long-term lifecycle management-specifically, the "Sunset" or Decommissioning stage of AI systems. In my practical experience, AI models are not static assets; their performance, safety, and ethical alignment degrade over time due to "data drift" or shifting societal norms. Therefore, I believe Dynamic Post-Market Oversight is a critical emerging issue. Two specific dimensions need to be addressed: 1. Mandatory Re-evaluation Triggers: Governance frameworks must define specific conditions (e.g., a significant shift in demographic data or a threshold of performance decay) that trigger a mandatory "re-inventory" and safety re-certification of a live model. 2. AI Decommissioning (Sunset) Protocols: Currently, there is no global consensus on how to safely "retire" an AI system. When a model is no longer compliant or reliable, we need interoperable protocols for its controlled withdrawal to prevent residual harms or "zombie models" from continuing to influence critical decision-making processes. By integrating Lifecycle Expiry into AI governance, we move from a "one-time compliance" mindset to a "continuous accountability" model. In this model, AI governance becomes a lifecycle-bound system of continuous accountability, where AI systems remain subject to ongoing re-evaluation and structured retirement as their operational relevance changes.

In enterprise environments where AI systems are deployed across multiple functions such as R&D, human resources, and supply chain, existing governance gaps are having a direct and practical impact on how AI is developed, scaled, and trusted. One of the most significant challenges is the lack of consistent governance structures across domains. Although many organizations have introduced AI governance practices, these are often implemented in a fragmented manner at the team or function level. As a result, AI systems are described and assessed using inconsistent definitions, making it difficult to compare risk, ensure accountability, or scale governance approaches across different business areas. A second key challenge is the limited maturity of lifecycle governance, particularly beyond deployment. In practice, AI systems are often treated as static assets after release, despite continuous changes in data, usage context, and external conditions. This leads to issues such as model drift, outdated decision logic, and unclear responsibility for system retirement or decommissioning. A third challenge is the lack of measurable indicators for governance effectiveness. While governance mechanisms are widely implemented, organizations still struggle to evaluate whether these controls are actually improving safety, reliability, or decision quality in a systematic way. At the same time, there are emerging opportunities. The adoption of AI asset management practices, including model identification, traceability systems, and workflow-embedded governance checkpoints, is creating a foundation for more structured and operational governance approaches. These developments enable greater visibility and control over AI systems throughout their lifecycle. The most significant impact of current governance gaps is that AI systems are becoming widely deployed faster than governance structures can consistently define, measure, and manage them at scale.

The AI Dialogue can play a critical role in advancing international cooperation by shifting the focus of AI governance from high-level principles toward shared operational foundations that enable comparability and coordination across jurisdictions. Today, one of the main barriers to international cooperation is not the lack of agreement on values such as safety, trustworthiness, or accountability, but the absence of common structural building blocks for describing and governing AI systems in practice. As a result, even when principles are aligned, implementation remains fragmented across countries, sectors, and organizations. In this context, the AI Dialogue can serve three important functions. First, it can facilitate convergence on shared governance primitives, such as standardized ways of describing AI systems, their intended use, and associated risks. This would create a common "language layer" that enables different governance systems to become mutually interpretable. Second, it can support the development of lifecycle-based governance approaches, including post-deployment monitoring, continuous evaluation, and responsible decommissioning of AI systems. This would help move global discussions beyond deployment-focused frameworks toward full lifecycle accountability. Third, it can promote early alignment on measurable indicators of governance effectiveness, enabling countries and organizations to compare outcomes such as safety performance, transparency, and reliability in a more systematic way. Overall, the most important contribution of the AI Dialogue would be to enable a transition from principle alignment to structural and operational interoperability in AI governance, thereby making international cooperation not only possible at the policy level, but actionable in real-world implementation.

The AI Dialogue should build upon and connect existing international and multi-stakeholder initiatives that are already advancing important components of AI governance, while addressing a key gap: the lack of cross-framework interoperability and shared operational structure. At the international level, it can build on emerging regulatory and policy efforts such as the OECD AI Principles, the G7 Hiroshima AI Process, and the EU AI Act, which provide important foundations for trustworthy and risk-based AI governance. It can also connect with technical and standard-setting bodies such as ISO/IEC AI standards initiatives, which are developing foundational definitions and risk management approaches. In addition, it should engage with industry-led and open ecosystem initiatives, including open-source AI development communities and model evaluation efforts, which are contributing to transparency, benchmarking, and reproducibility in AI systems. These initiatives collectively reflect significant progress, but they remain partially fragmented across regions and sectors. The added value of the AI Dialogue would be to function as a coordination layer across these existing efforts, rather than introducing parallel frameworks. Specifically, it can help translate high-level principles and technical standards into shared operational governance building blocks that can be consistently applied across jurisdictions. This includes fostering convergence on (1) common ways of structurally describing AI systems, (2) lifecycle-based governance approaches that extend beyond deployment into monitoring and decommissioning, and (3) emerging practices for evaluating governance effectiveness in measurable terms. By doing so, the AI Dialogue can help bridge the gap between policy principles, technical standards, and enterprise implementation. Its unique contribution would be enabling interoperability across existing governance ecosystems, ensuring that different initiatives do not evolve in isolation but instead become mutually reinforcing components of a more coherent global AI governance architecture.

Different stakeholders can contribute to the AI Dialogue in ways that reflect their actual proximity to AI systems, rather than only their formal roles in policy discussions. Governments can share how AI governance is being interpreted and applied in real regulatory contexts, including where current rules are working and where they are still unclear or difficult to enforce. International organizations can help connect these experiences and highlight common patterns across countries. From my perspective, the most valuable input from industry is very practical: what actually breaks when AI governance is implemented in real workflows. This includes issues like inconsistent definitions across teams, governance becoming a "check-the-box" exercise, or systems that are hard to maintain once they are deployed at scale. These are often the gaps that don't appear in formal frameworks but decide whether governance actually works. The technical community can contribute by making governance more usable in practice—through better standards for describing AI systems, tracking them, and evaluating them consistently. Civil society and academia can help keep attention on the longer-term social and ethical consequences that may not be visible in day-to-day implementation. In terms of format, I think the Dialogue would be more effective if it is not only a discussion space, but also a structured learning system. For example, inputs could be organized around real use cases, with clear separation between policy intent, implementation reality, and technical constraints. Over time, these inputs could be synthesized into shared reference patterns that different stakeholders can actually reuse. The value of the Dialogue, in the end, will depend on whether it helps translate diverse experiences into something that is not just discussed globally, but can also be applied locally in a consistent way.

In global discussions on AI governance, some of the most underrepresented perspectives are from groups that are highly affected by AI systems, but have limited access to decision-making spaces where these systems are defined and regulated. This includes workers in operational roles where AI is increasingly used to make or support decisions, such as recruitment, logistics, customer service, and content moderation. It also includes smaller companies and organizations in developing contexts that adopt AI tools but do not have the capacity to shape how those tools are governed or evaluated. In addition, individuals whose data is continuously used by AI systems are often not meaningfully represented in governance discussions. What is often missing is not awareness of these groups, but practical mechanisms to bring their experience into structured governance inputs. To address this, inclusion should go beyond open consultation formats. One approach is to anchor discussions in real use cases, where stakeholders are invited to respond to concrete system scenarios rather than abstract principles. Another approach is to involve intermediary institutions—such as industry associations, labor organizations, and local technology partners—that can translate lived experience into actionable governance insights. In addition, feedback mechanisms from deployed AI systems themselves—such as incident reporting, user override patterns, and system performance in real environments—can serve as indirect but important signals of underrepresented perspectives. Ultimately, inclusion in AI governance should not only be about representation in discussions, but about ensuring that the effects of AI systems are systematically reflected back into how those systems are governed.

To foster more meaningful and dynamic engagement, the AI Dialogue could benefit from formats that move beyond traditional consultations and instead connect discussion directly to real-world AI systems and implementation challenges. One effective format would be use-case-based dialogue sessions, where stakeholders are not only asked to express opinions on abstract principles, but to engage with concrete AI scenarios drawn from sectors such as healthcare, hiring, finance, or public services. This would help surface practical governance tensions that are often hidden in high-level discussions. Another useful approach would be multi-layered structured contributions, where inputs are explicitly separated into three levels: policy intent, operational constraints, and technical implementation realities. This would make it easier to compare perspectives across stakeholders who often speak different "languages" when discussing AI governance. In addition, the Dialogue could incorporate iterative synthesis cycles, where inputs are not only collected once, but periodically consolidated into evolving shared reference outputs. These could include living frameworks, governance patterns, or common vocabulary for describing AI systems. A further innovation could be feedback-loop mechanisms from deployed AI systems, where real-world signals—such as incident reports, user overrides, or system performance drift—are periodically fed back into discussions. This would ensure that engagement is grounded in how AI behaves in practice, not only how it is designed in theory. Finally, hybrid formats combining in-person dialogues with structured digital collaboration platforms could help sustain engagement over time, allowing stakeholders to contribute continuously rather than only during formal events. The most effective engagement formats are those that make AI governance discussions more situated, iterative, and connected to real system behavior, rather than purely conceptual exchanges.

From what I've seen in real enterprise settings, effective AI governance is less about having a single perfect framework, and more about whether governance can actually "stick" inside day-to-day work. At the policy level, frameworks like risk-based regulation (for example, the EU AI Act approach) are useful because they force organizations to think in terms of different levels of impact and responsibility, rather than treating all AI systems the same. Principles like those from OECD also help create a shared baseline, even if they are not directly operational. In practice, the most useful governance approaches tend to be very concrete. For example, some organizations assign unique IDs to AI models and datasets so they can actually track what is running in production. Others embed simple checkpoints into existing development workflows, so governance is not an external review step, but part of the release process itself. These kinds of mechanisms matter more than formal documentation because they influence real behavior. On the tooling side, what really helps are systems that make AI visible over time-things like tracking model versions, monitoring performance drift, and linking outputs back to training data. Without this kind of visibility, governance quickly becomes theoretical after deployment. Another thing that works better than expected is standardizing how AI systems are described internally. When different teams use a shared structure to explain what their model does, who owns it, and where it is used, it becomes much easier to compare systems and spot risks early. This highlight the approaches that work best are not the most complex ones, but the ones that make governance part of how work actually happens, rather than something added on top.