NeoMundi Research

A successful outcome would be the recognition that effective AI governance requires operational measurement during system execution, not only principles or ex-post evaluation. The Dialogue could establish the need for shared, verifiable and provider-independent measurement frameworks enabling real-time oversight, traceability and risk monitoring. Concretely, this could translate into: – the identification of measurement as a foundational layer of AI governance, – the encouragement of interoperable metrics applicable across systems and jurisdictions, – and the integration of such measurement capabilities into existing regulatory and technical frameworks. This would mark a shift from declarative governance to operational governance, making oversight technically practicable and internationally comparable.

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

These priorities are closely linked: effective human oversight, safety, transparency and interoperability all depend on the ability to observe AI systems while they operate. Today, most governance approaches remain declarative or rely on post-hoc evaluation. Without a continuous and verifiable measurement layer, oversight remains partial and difficult to operationalize in practice. Measurement frameworks that are independent of providers and applicable across architectures can serve as a common reference between actors and jurisdictions. They enable comparable risk monitoring, auditable traceability and technically grounded supervision. At the same time, such approaches can support capacity-building by providing lightweight, deployable tools that do not require extensive infrastructure or access to proprietary systems. These elements together form the basis for a governance model that is not only principled, but operational.

A key transversal gap is the absence of a shared metrology for AI systems. Current discussions focus on principles, standards and risk categories, but lack a common measurement layer enabling continuous and comparable observation of system behavior in real conditions. Without such a metrology, governance remains fragmented: oversight is difficult to operationalize, interoperability is limited, and accountability relies on indirect signals. Developing and recognizing measurement infrastructures as a foundational layer of AI governance could address this gap, enabling coordination across technical, regulatory and geopolitical dimensions.

In the current landscape, the absence of a shared and continuous measurement layer for AI systems creates several structural limitations across regions and sectors. In Europe, and more broadly in regulated environments, governance frameworks are advancing rapidly, but their operationalization remains challenging. Oversight mechanisms often rely on documentation, testing phases or ex-post audits, which do not fully capture system behavior in real conditions. This creates a gap between regulatory intent and technical enforceability. For organizations and public actors, this results in uncertainty: risk is difficult to monitor continuously, accountability remains indirect, and interoperability between systems or jurisdictions is limited by the lack of a common reference. In emerging and resource-constrained environments, these limitations are amplified. The absence of lightweight, provider-independent tools restricts the ability to implement meaningful oversight without relying on external actors or proprietary infrastructures. At the same time, recent advances in AI systems increase both capabilities and systemic risks, making continuous observation more critical. The emergence of measurement-based approaches represents a key opportunity. By enabling real-time, verifiable and comparable observation of AI systems, they can reduce the gap between governance frameworks and operational reality. This would support more effective oversight, facilitate interoperability across jurisdictions, and strengthen capacity-building by providing deployable and scalable tools for a wide range of actors.

The Dialogue can play a key role by shifting international cooperation on AI governance from alignment of principles to alignment of operational capabilities. Today, cooperation is often limited by the absence of shared technical references. While principles may converge, their implementation remains fragmented across jurisdictions, systems and actors. The Dialogue could support the emergence of common measurement approaches that enable continuous, verifiable and comparable observation of AI systems in real conditions. Such approaches can function as a neutral technical layer, independent of specific providers or architectures. By promoting interoperable measurement frameworks, the Dialogue can facilitate coordination between regulatory regimes, reduce asymmetries between actors, and strengthen mutual trust. This would enable a form of cooperation grounded not only in shared values, but in shared observability, making governance more consistent, auditable and actionable across borders.

The Dialogue could build on and connect with existing international efforts on AI governance, including regulatory frameworks, standardization initiatives and capacity-building programs. Its added value would be to introduce and coordinate a measurement-oriented layer across these initiatives. While many current efforts focus on principles, risk categories or model capabilities, fewer address how systems can be continuously observed and compared in operational settings. The Dialogue could act as a convening platform to: – encourage the development of interoperable measurement frameworks, – support open scientific publication and reproducibility of methodologies, – and facilitate the integration of such approaches into regulatory and technical ecosystems. This would complement existing initiatives by providing a common operational reference, enabling greater coherence between standards, policies and real-world system behavior. In this sense, the Dialogue would not duplicate existing efforts, but strengthen them through a shared layer of measurement and observability.

Different stakeholders can contribute most effectively if the Dialogue is structured around both principles and operational tools. Governments can provide regulatory frameworks and ensure alignment with public policy objectives. Technical communities and researchers can contribute measurement methodologies, validation protocols and open scientific work. Private sector actors can test and integrate these approaches in real-world systems. Civil society can assess their societal implications and ensure accountability. To enable meaningful participation, the Dialogue could combine high-level discussions with technical working tracks focused on concrete outputs, such as shared measurement approaches or interoperability mechanisms. A structured interface between policy and technical layers would allow contributions to be cumulative, comparable and actionable, rather than purely declarative.

Several perspectives remain underrepresented in global AI governance discussions, particularly actors from resource-constrained environments, smaller institutions and technical communities working outside major platforms. These actors often lack access to proprietary systems, large-scale infrastructure or internal evaluation capabilities, which limits their ability to meaningfully participate in governance processes. One way to strengthen inclusion is to promote approaches that do not depend on access to underlying models or sensitive data, but instead rely on observable system behavior. Measurement-based frameworks that are provider-independent and lightweight can enable broader participation by lowering technical and economic barriers. By focusing on shared, observable metrics, the Dialogue can create a more level playing field, allowing a wider range of actors to contribute to governance discussions with comparable evidence.

To foster meaningful and dynamic engagement, the Dialogue could integrate formats that combine policy discussion with empirical and technical exploration. This could include: – collaborative technical workshops focused on concrete governance tools and methodologies, – shared evaluation exercises using common measurement frameworks, – and open repositories of reproducible approaches and results. Such formats would allow participants not only to exchange views, but to work on comparable and testable elements, strengthening mutual understanding. Introducing a layer of shared observability into these formats would enable discussions to be grounded in comparable evidence, making engagement more concrete, iterative and outcome-oriented. This would complement traditional dialogue formats by adding a practical dimension, bridging the gap between principles and implementation.

Several emerging approaches contribute to more effective AI governance by moving from principles toward operational practices. Regulatory frameworks increasingly emphasize accountability, transparency and human oversight, while technical practices such as red teaming, benchmarking and auditing aim to evaluate system behavior. However, these approaches often remain partial, as they focus on specific phases of the system lifecycle rather than continuous observation in real conditions. A complementary direction is the development of measurement-based approaches that enable real-time, verifiable and comparable observation of AI systems during operation. Such approaches can provide a continuous signal of system behavior, supporting oversight, traceability and risk monitoring. In practice, this can take the form of measurement infrastructures that operate independently of model providers and without requiring access to underlying data or content. These infrastructures can be integrated into existing systems to provide continuous monitoring and auditable outputs. In parallel, open scientific publication, shared evaluation datasets and reproducible methodologies are essential to ensure trust, comparability and broad participation. Together, these elements point toward a governance model that combines regulatory frameworks, technical evaluation and continuous measurement, enabling more effective and operational oversight of AI systems.