SEILX

A successful outcome would move AI governance from principles to enforceable infrastructure. Today, most frameworks define what should happen, but lack mechanisms to verify what actually happened in practice. This creates a gap between policy and reality, especially in high-risk use cases. The first Global Dialogue should therefore establish a shared direction around verifiable accountability, including: – Agreement that high-risk AI systems must produce auditable, tamper-resistant records of decision events – A roadmap for implementing independent verification mechanisms, not controlled solely by system operators – Alignment across jurisdictions on minimum standards for evidence integrity and traceability – Pilot initiatives that demonstrate how such systems can be implemented in real-world environments Without verifiability, transparency and accountability remain interpretative rather than enforceable. A successful dialogue would not only define principles, but initiate the transition toward infrastructure that makes those principles provable in practice.

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

The selected priorities are closely connected and depend on a shared foundation: the ability to verify what AI systems actually do. "Safe, secure and trustworthy AI" cannot be achieved without mechanisms to validate system behavior at the moment of execution. Similarly, "transparency and accountability" require more than post-hoc explanations - they require verifiable records. The protection of human rights increasingly depends on whether individuals can prove that harm occurred, particularly in cases involving automated decisions or AI-generated content. Finally, interoperability of governance approaches is critical. Without aligned standards for evidence, traceability, and verification, different jurisdictions will interpret and enforce AI-related incidents inconsistently. Together, these priorities point to the same underlying need: a shift from policy-based governance to evidence-based governance. This is where meaningful accountability becomes possible.

A critical cross-cutting issue not sufficiently captured is the absence of verifiable evidence infrastructure for AI systems. Current discussions address transparency, accountability, and safety, but often assume that system behavior can be reliably reconstructed after the fact. In practice, this is rarely the case. Logs can be incomplete, altered, or controlled by the same entities responsible for the system. This creates structural ambiguity in determining what actually occurred during an AI-driven event. This issue becomes especially acute in high-risk contexts such as child safety, financial decisions, and digital harm, where outcomes depend on whether evidence is trusted and admissible. Without mechanisms to: - capture decision events at the point of execution - secure them in a tamper-resistant manner - and allow independent verification governance frameworks risk becoming declarative rather than enforceable. Addressing this gap should be treated as a foundational requirement for future AI governance.

In Western Europe, including the EU, AI governance is advancing rapidly through frameworks such as the AI Act. This creates clarity at a regulatory level, but also exposes a critical gap at the implementation level. The most significant challenge is the lack of mechanisms to verify AI system behavior in practice. Organizations are increasingly required to assess, document, and justify AI-driven decisions. However, in many real-world deployments, it remains difficult to establish: – what exact input triggered a decision – how the system processed it at that moment – whether the output has been altered or contested after the fact This creates operational ambiguity. Compliance becomes documentation-heavy but evidence-light. In high-risk scenarios — including financial decisions, automated moderation, and cases involving digital harm — outcomes often depend on interpretation rather than verifiable facts. At the same time, this gap represents a major opportunity. There is growing demand for infrastructure that enables: – real-time capture of decision events – tamper-resistant storage – independent verification across stakeholders Such capabilities would strengthen trust, improve enforceability, and reduce legal uncertainty across jurisdictions. For the EU in particular, aligning regulatory ambition with technical verifiability will be critical. Without this, there is a risk that governance frameworks become difficult to operationalize consistently across sectors. Bridging this gap could position the region as a global leader not only in AI regulation, but in enforceable, evidence-based AI governance.

The AI Dialogue can play a critical role by shifting international cooperation from alignment of principles to alignment of enforceable mechanisms. Today, many jurisdictions agree on high-level goals such as transparency, accountability, and safety. However, cooperation breaks down in practice because there are no shared technical standards for verifying AI system behavior. As a result, the same incident can be interpreted differently across countries, depending on available evidence, local regulations, and trust in system operators. The Dialogue can address this by: – Promoting global minimum standards for verifiable AI decision records – Supporting interoperable approaches to evidence integrity and traceability – Encouraging the development of independent verification frameworks that can be recognized across jurisdictions – Facilitating cross-border pilots to test how such systems function in real-world scenarios Without this layer, international cooperation risks remaining declarative rather than operational. By focusing on verifiability as a shared foundation, the AI Dialogue can enable more consistent enforcement, reduce legal fragmentation, and build trust across borders. This would mark a shift from coordinating policies to coordinating enforceable reality.

The AI Dialogue should build upon existing initiatives such as the EU AI Act, OECD AI Principles, UNESCO's Recommendation on AI Ethics, and emerging national regulatory frameworks. These efforts have successfully established shared values and risk-based approaches. However, they primarily focus on governance principles, compliance processes, and oversight structures. A key gap across these initiatives is the lack of standardized mechanisms for verifying AI system behavior at the moment of execution. The added value of the AI Dialogue would be to: – Connect policy frameworks with technical implementation standards – Promote the development of verifiable evidence infrastructure as a common layer across jurisdictions – Encourage alignment on how AI-related incidents are documented, secured, and independently validated – Support collaboration between regulators, technical communities, and system builders to define these standards By doing so, the Dialogue can move beyond reinforcing existing principles and instead address the operational gap that limits their effectiveness. This would complement current initiatives by making their objectives measurable, enforceable, and consistent across borders.

Different stakeholders should not only contribute perspectives, but be assigned clearly defined roles in a structured, outcome-driven process. Governments should define regulatory objectives and constraints, but also participate in testing how these rules function under real-world conditions. The private sector should provide access to real systems, data flows, and deployment scenarios, enabling practical evaluation rather than theoretical discussion. The technical community should design and validate mechanisms for verification, traceability, and system integrity, ensuring that governance is technically enforceable. Civil society should contribute real-world cases, impact assessments, and challenge assumptions, ensuring that governance reflects lived consequences. Academia should support with independent validation, methodology, and long-term research perspectives. To structure this effectively, the AI Dialogue should be organized into: – Thematic execution tracks (e.g. safety, accountability, interoperability) – Cross-stakeholder working groups with mandatory deliverables – Defined output formats (standards drafts, audit frameworks, test protocols) – Public documentation of results, including failures and limitations Participation should be tied to contribution: stakeholders are expected to bring data, systems, or validated insights—not only opinions. This structure ensures that the Dialogue produces usable outcomes, not just alignment in language.

Several critical perspectives remain underrepresented in global AI governance discussions—particularly those closest to real-world system failures. These include: – Individuals directly impacted by AI decisions, especially in high-stakes domains such as finance, healthcare, law enforcement, and social services – Practitioners responsible for investigating failures, including auditors, incident responders, and forensic analysts – Operators and integrators of AI systems, who understand how systems behave in production rather than in controlled environments – Smaller organizations and non-dominant markets, where resources for compliance and oversight are limited A key gap is that current discussions are dominated by policy, research, and large technology actors, while those dealing with consequences and system breakdowns are less visible. To address this, inclusion should be structured, not symbolic: – Require real case submissions from affected individuals and organizations – Integrate failure analysis sessions where incidents are examined in detail – Provide secure channels for sharing sensitive or high-risk cases – Ensure representation in working groups is tied to operational experience, not only institutional affiliation Without these perspectives, governance risks being designed around assumptions rather than reality. Including those who experience, investigate, and manage AI failures is essential to building systems that are accountable in practice.

To foster meaningful and dynamic engagement, the AI Dialogue should move beyond discussion-based formats and incorporate execution-oriented structures. Current formats such as panels and roundtables are valuable for perspective sharing, but they rarely produce verifiable or actionable outcomes. More effective formats would include: – Scenario-based simulations: Participants work through real-world AI incidents (e.g. failed decisions, harm cases) and are required to demonstrate how their governance frameworks perform in practice – Live system audits: Organizations present actual AI systems and allow structured, multi-stakeholder review of decision processes, risks, and evidence availability – Cross-border pilot programs: Small-scale implementations where multiple jurisdictions test shared governance approaches under real conditions – Technical sandboxes: Collaborative environments where regulators, developers, and auditors co-design and test verification, accountability, and safety mechanisms – Outcome-driven working groups: Each group is tasked with delivering a concrete output (e.g. a draft standard, validation method, or interoperability protocol) rather than general recommendations Additionally, all sessions should be structured around measurable outputs, with clear documentation of what was tested, what failed, and what worked. Without this shift, engagement risks remaining conceptual. With it, the Dialogue can become a platform for building and validating the foundations of real-world AI governance.

Several existing approaches contribute meaningfully to AI governance, but most remain incomplete without mechanisms for verifiable execution. Examples of strong foundations include: - The EU AI Act, which establishes a risk-based framework and introduces obligations for high-risk systems - NIST AI Risk Management Framework, providing structured guidance on identifying, assessing, and managing AI risks - OECD AI Principles, which define widely accepted values such as transparency, accountability, and human oversight - Model cards and system documentation practices, improving transparency around system behavior and limitations These initiatives are important, but they primarily define intent, process, and documentation-rather than enforceable verification. A critical next step is to complement these with technical mechanisms that ensure claims can be validated in practice. Promising approaches include: - Tamper-resistant logging of AI decisions and system states - Cryptographic sealing of events to ensure integrity over time - Independent audit layers capable of reconstructing decision processes - Standardized evidence schemas for cross-system and cross-border verification Governance frameworks should evolve from "declared compliance" to provable compliance, where organizations can demonstrate-not just describe-how decisions were made and whether systems behaved as intended. The combination of policy frameworks with verifiable technical infrastructure is what will enable AI governance to function reliably at scale.