Independent Researcher in AI Governance | Normative Theory & Philosophy of Law | IHL & International Law | Accountability Gap in AI Systems

A successful first Global Dialogue on AI Governance would produce outcomes that move beyond general principles toward shared, operational clarity. First, it should establish a common conceptual baseline for accountability. Current discussions often rely on compliance as the primary indicator of accountability. The Dialogue would benefit from recognizing a structural distinction between compliance—whether a rule applies to a system—and directed responsibility—whether a specific agent can answer for a specific decision. Clarifying this distinction would reduce ambiguity across governance approaches and support more precise policy design. Second, the Dialogue should identify a minimal set of interoperable governance concepts that can travel across legal systems and institutional contexts. A shared vocabulary for key terms such as accountability, human oversight, and transparency would enhance coherence and facilitate mutual understanding among States and stakeholders. Third, it should generate practical guidance for high-stakes AI systems. This includes encouraging dual-level assessment of compliance and decision-level answerability, and exploring the role of answerability thresholds as a condition for deployment in contexts where decisions have significant consequences. Finally, success would involve embedding these insights into ongoing processes, including capacity-building and future Dialogues, ensuring that the outcomes are not only discussed but can be progressively operationalized within diverse governance frameworks. Such outcomes would strengthen transparency, human oversight, and the overall coherence of international AI governance efforts.

Transparency, accountability, and human oversight

Title: From Compliance to Answerability: A Structural Test for High-Stakes AI Governance Summary: Current AI governance frameworks focus on compliance: whether a rule applies to a system and is satisfied. This is necessary, but insufficient for accountability. A distinct dimension is often left unarticulated. Directed responsibility concerns whether a specific agent can answer for a specific decision by providing reasons for why that outcome was produced rather than relevant alternatives. Compliance and directed responsibility are different in kind, not degree. No increase in compliance mechanisms can produce directed responsibility. As a result, systems may be fully compliant while no actor can substantively account for particular outcomes. This condition constitutes an accountability gap. It is not merely a failure of enforcement, but a structural feature of certain decision architectures. This submission proposes recognizing answerability as a distinct governance requirement alongside compliance in high-stakes AI systems. Policy Recommendations: 1. Dual-level assessment Governance should independently evaluate compliance and directed responsibility. A system is not accountable solely because it is compliant. 2. Answerability threshold Deployment in high-stakes contexts should require at least one identifiable agent capable of explaining specific decisions and their selection over alternatives. 3. Accountability gap as risk The accountability gap should be treated as an independent governance risk, supported by shared audit vocabulary, reporting practices, and integration into interoperability and capacity-building efforts.

Yes. A key cross-cutting issue not explicitly captured by the listed themes is the structural distinction between compliance and decision-level answerability. Current thematic areas-such as transparency, accountability, human oversight, and interoperability-presume that accountability can be strengthened through improved compliance mechanisms. However, this overlooks a distinct question: whether a specific agent can substantively answer for a specific decision. Compliance evaluates whether rules apply to systems and are satisfied. It operates at a positional, system-level. By contrast, answerability concerns whether an identifiable agent can provide reasons for why a particular outcome occurred rather than relevant alternatives. It operates at the level of individual decisions. This distinction cuts across all thematic areas. In "safe and trustworthy AI," systems may meet all requirements yet still produce outcomes no actor can explain. In "human oversight," the presence of a human in the loop does not guarantee that the human can answer for the decision. In "interoperability," governance frameworks may align on compliance standards while remaining incompatible in how responsibility is attributed. The resulting condition-where compliance is satisfied but no agent can answer for specific outcomes-constitutes an accountability gap. This is not a failure of implementation but a structural feature of certain AI-enabled decision architectures. Recognizing answerability as a distinct governance requirement would strengthen existing themes, provide a shared operational vocabulary, and improve the coherence of international AI governance efforts.

In our sector, the central governance gap concerns the divergence between compliance and decision-level answerability. Significant advances have been made in developing compliance-oriented frameworks: procurement rules, audit procedures, documentation standards, and risk classification schemes. These have improved formal oversight and regulatory alignment. However, they primarily operate at the level of system classification and rule applicability. The challenge arises at the level of individual decisions. In practice, AI-enabled decision processes—whether in administrative allocation, risk assessment, or operational support—can satisfy all applicable compliance requirements while leaving no identifiable agent able to provide a substantive explanation for a specific outcome. This creates an accountability gap: decisions are produced, actions are taken, but answerability is diffused or absent. This gap has two implications. First, it limits the effectiveness of human oversight, which often becomes supervisory in form but not substantive in function. Second, it creates institutional exposure, as responsibility cannot be clearly attributed when decisions are contested. At the same time, there is a significant opportunity. Recognizing answerability as a distinct governance requirement would allow institutions to move beyond procedural compliance toward decision-level accountability. This would strengthen public trust, improve auditability, and enhance the coherence of governance frameworks across jurisdictions. Addressing this gap requires not more compliance, but a complementary focus on who can answer for specific decisions and on what basis.

The AI Dialogue can play a practical convening role in international cooperation by turning broad agreement into shared operational understanding. As established by General Assembly resolution 79/325, the Dialogue is a multi-stakeholder platform for Governments and relevant stakeholders to discuss international cooperation, share best practices and lessons learned, and facilitate open, transparent, and inclusive discussions on AI governance. It is also designed to feed into the annual report of the Scientific Panel and to support follow-up through recurring Dialogue sessions. Its most important contribution is to help build coherence across a fragmented governance landscape. The draft structure frames this explicitly: the Dialogue can connect existing initiatives, identify practical lessons, and promote greater coherence, interoperability, and mutual understanding across different processes. For international cooperation to advance meaningfully, the Dialogue should do more than restate principles. It should help participants develop a common conceptual and policy vocabulary, including on transparency, accountability, and human oversight, so that governance approaches can be compared and aligned across contexts. It should also surface concrete capacity gaps and implementation challenges, especially for developing countries, and identify where technical cooperation is needed. In short, the Dialogue can serve as an inclusive United Nations forum that transforms abstract consensus into shared methods, practical cooperation, and progressively interoperable governance arrangements.

The AI Dialogue should build upon existing United Nations and multi-stakeholder mechanisms already operating in the AI governance space, especially the Independent International Scientific Panel on AI, the broader Global Digital Compact architecture, and the connected processes around the ITU AI for Good Global Summit and other relevant UN conferences and meetings. It should also connect with ongoing work by Member States, international organizations, the technical community, academia, civil society, and standards development bodies, all of which are explicitly contemplated in the Dialogue's mandate and draft structure. Its added value is not to duplicate those efforts, but to provide a single inclusive UN venue where their outputs can be compared, harmonized where possible, and translated into common governance vocabulary and practical cooperation. In particular, the Dialogue can help identify points of convergence across different governance approaches, surface implementation gaps, and support interoperability without forcing uniformity. It can also strengthen capacity-building by connecting normative discussion with real institutional and technical needs, especially for developing countries. In my view, its most important contribution is to move the conversation from parallel initiatives to a shared operational understanding of accountability, transparency, and human oversight in AI governance. That is where the Dialogue can create genuine added value.

Different stakeholders can contribute in complementary ways. Governments should set priorities, identify regulatory and capacity gaps, and signal where international cooperation is most needed. The Independent International Scientific Panel should provide evidence-based framing and help distinguish what is technically feasible from what requires policy choice. The private sector, civil society, academia, the technical community, and international organizations should contribute use cases, implementation experience, risk perspectives, and practical proposals for interoperability, transparency, accountability, and human oversight. This broad participation is exactly what the mandate and draft note envisage. The Dialogue should be structured to convert this diversity into usable outputs. A short high-level governmental segment should set priorities, followed by a multi-stakeholder segment to surface practical experiences, then thematic breakouts focused on the four mandated clusters. The Panel's annual report should be presented early enough to inform discussion, not merely as a ceremonial item. Thematic discussions should be moderated, time-bounded, and oriented toward concrete policy questions rather than general statements. The added value of the Dialogue lies in synthesis: connecting existing initiatives, identifying points of convergence, and translating dispersed work into shared language and practical cooperation. To achieve that, the format should privilege focused interventions, balanced speaking time, co-chair summaries, and a final closing session that captures priority actions and areas of agreement for follow-up.

Independent researchers are among the most underrepresented voices in global discussions on AI governance. Unlike institutional actors, they often operate outside formal organizational structures, funding pipelines, and diplomatic channels, yet they frequently contribute original conceptual and interdisciplinary work that does not emerge from established policy or industry frameworks. Current processes tend to privilege States, large organizations, and well-resourced stakeholders. While this is understandable, it risks narrowing the range of perspectives, particularly at the level of foundational concepts such as accountability, human oversight, and responsibility. Independent researchers can play a distinct role here by introducing structurally new approaches, critical distinctions, and cross-disciplinary insights that may not align with existing institutional agendas. To include these voices more effectively, the Dialogue could formalize accessible entry points that do not depend on institutional affiliation. This may include open submission tracks with transparent selection criteria, recognition of independent expertise alongside organizational representation, and structured opportunities to contribute to thematic discussions and background materials. In addition, capacity-building efforts could explicitly extend to independent contributors, particularly in regions where institutional support for research is limited. Ensuring that participation mechanisms are inclusive of individuals, not only organizations, would strengthen the diversity and depth of the Dialogue. Incorporating independent researchers is not only a matter of representation, but of epistemic completeness. Broadening participation in this way would enhance the Dialogue's ability to identify gaps, challenge assumptions, and develop more robust and adaptable approaches to AI governance.

The most effective formats would be those that preserve the Dialogue's multi-stakeholder structure while forcing substantive exchange rather than parallel monologues. The draft structure already points in the right direction: high-level plenaries, thematic breakouts, moderated exchanges, and sessions designed to surface practical policy and cooperation insights for the Co-Chairs' summary. In my view, three formats would add the most value. First, short scenario-based breakouts. Instead of open-ended statements, participants could respond to a concrete governance problem, compare approaches, and identify where interoperability breaks down. This would turn abstract principles into operational discussion. Second, mixed-stakeholder problem-solving roundtables. Bringing together Governments, the scientific panel, industry, civil society, and technical experts in small moderated groups would make it easier to compare lived implementation experience with normative and technical perspectives. The Dialogue's mandate already supports this kind of inclusive exchange. Third, a live synthesis format at the end of each thematic block. A rapporteur or co-chair could capture convergences, divergences, and unresolved questions in real time, so that discussion feeds directly into the final summary rather than remaining fragmented. This would reinforce the Dialogue's objective of coherence across the broader AI governance landscape. I would also recommend strict time limits for interventions, guided questions in advance, and a format that privileges specific policy asks over general statements. That would make the Dialogue more dynamic, more balanced, and more likely to generate usable outcomes.

Examples that are already moving in the right direction are those that turn "responsible AI" into operational requirements rather than slogans. First, the Global Digital Compact is a strong policy model because it treats AI governance as a lifecycle issue and explicitly links it to transparency, accountability, robust human oversight, interoperability, and capacity-building. That is a useful template because it pushes governance beyond general principles and toward implementable obligations. Second, General Assembly resolution 79/325 is itself an important platform model: it creates the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance as a recurring multistakeholder mechanism for sharing evidence, lessons learned, and policy-relevant convergence. That kind of platform matters because it can connect fragmented initiatives and produce shared vocabulary across jurisdictions. Third, the draft structure for the Dialogue is a practical example of a useful format: high-level plenaries, thematic breakouts, and moderated multi-stakeholder exchanges. This is valuable because it forces discussion into concrete policy clusters-capacity gaps, safe and trustworthy AI, interoperability, and human rights-rather than leaving it at the level of abstract principle. In my view, the most effective approach is one that combines these elements: a lifecycle-based policy framework, an evidence-producing scientific panel, and a recurring multistakeholder forum that can translate broad consensus into interoperable governance practices. The added value is not more rhetoric; it is shared operational clarity.