Digital Childhood Council

The first Global Dialogue on AI Governance would mark a genuine success if it moves the international community from aspirational principles toward structural accountability and establishing shared understanding of what governance requires at the level of system architecture, not only at the level of policy. Meaningful success would include agreement on three foundational conditions. First, that accountability requires more than attribution after harm occurs. Governance frameworks that identify responsible parties only after exposure has already reached affected populations are reactive instruments, not accountability structures. A successful Dialogue would establish that pre-deployment binding including formal assignment of responsibility before systems interact with populations, is a minimum condition of governed deployment, not an aspirational standard. Second, that the feedback loop between identified harm and required behavioral change must be enforceable, not discretionary. Current deployment structures allow harm to be identified, reported, and documented without producing any binding obligation on the part of system developers to modify future behavior. Where financial consequences attach to deploying institutions or insurers rather than to the systems that generated the harm, the behavioral constraint function of accountability is structurally absent. A successful Dialogue would recognize this gap as a governance condition requiring specific remedy, not a contracting detail to be resolved bilaterally. Third, that correlated harm, such as where the same system produces similar harmful outputs across large populations simultaneously, requires governance frameworks designed for systemic risk, not individual incident response. Standard accountability mechanisms assume independent loss events. AI systems operating at scale defeat that assumption. International governance frameworks must reflect the actuarial and legal implications of correlated exposure. Success is not a declaration of principles. It is the beginning of a shared framework in which responsibility is formally attached before influence occurs, enforceable after harm is identified, and structurally connected to the systems that generate it.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights
• Social, economic, ethical, cultural, linguistic and technical implications of AI

These four priorities reflect a single underlying concern: that current AI governance frameworks assume accountability conditions that generative AI systems, as currently designed and deployed, do not structurally satisfy. Transparency, accountability, and human oversight is the primary priority because it addresses the most consequential gap in current governance. The absence of an enforceable mechanism connecting identified harm to required behavioral change. Where systems can cause harm, receive reports of that harm, and continue operating without binding obligation to modify future behavior, oversight is procedural rather than structural. Transparency without enforceability does not constitute accountability. Safe, secure and trustworthy AI is urgent because trustworthiness cannot be assessed solely by technical performance benchmarks. A system that produces correlated harm across large populations simultaneously where the same failure mode reaches millions of users before identification and correction can occur, presents systemic risk that standard safety frameworks were not designed to address. Trustworthiness requires that identified failures produce binding corrections, not discretionary responses. Protection and promotion of human rights follows directly from there. Where financial consequences for AI-generated harm attach to deploying institutions rather than to the systems that generated the harm, affected populations bear consequences while the systems that produced them are not structurally constrained from recurrence. This asymmetry of harm externalized to populations, consequence not internalized by systems is a human rights condition requiring governance remedy at the international level. Social, economic, ethical and cultural implications encompass the systemic effects of deploying accountability frameworks that are descriptive rather than operational and where responsibility is assigned in principle but cannot be enforced in practice. The long-term social and institutional consequences of normalized ungoverned influence at scale warrant urgent and sustained international attention.

Yes, the listed themes collectively address what governance frameworks should require of AI systems. A critical cross-cutting issue not captured by any of them is whether the structural conditions exist for those requirements to function, and for one category of system, they do not. Conventional accountability frameworks across safety, human rights, transparency, and oversight all share a foundational assumption: that when a system causes harm, the consequence of that harm reaches the entity responsible for the system's behavior in a way that produces a binding constraint on future conduct. This assumption holds for most institutional risk contexts but it does not hold for generative AI systems as currently designed and deployed. Generative AI systems do not contain internal mechanisms by which harm to an affected population re-enters the system as a corrective signal. Harmful outputs exit the system without generating any internal consequence. Financial liability, where it attaches at all, typically attaches to deploying institutions rather than to system developers and even where it reaches developers, the transmission mechanism that would convert financial consequence into behavioral correction is indirect, discretionary, and not governed by any enforceable international standard. The result is a structural condition in which accountability frameworks can assign responsibility, require transparency, and mandate oversight and still produce no binding constraint on future system behavior following identified harm. This is not a gap in policy; it is an architectural condition of current systems that policy frameworks were not designed to address. Any governance framework that does not specifically address the mechanism by which identified harm produces enforceable behavioral change in AI systems will remain descriptive rather than operational, regardless of how comprehensively it addresses the thematic areas listed above. This structural feedback gap is the cross-cutting issue the Dialogue cannot afford to leave unaddressed.

The governance gaps identified in my previous responses are already producing measurable challenges in institutional deployment contexts, particularly in education, insurance, and public sector settings where accountability frameworks are legally defined and non-delegable. In these sectors, institutions are deploying generative AI systems under governance assumptions that the systems do not structurally satisfy. Responsibility is assigned to deploying institutions through existing duty of care and liability frameworks. The systems generating outputs that trigger that responsibility contain no internal mechanism by which harm produces a binding constraint on future behavior. The financial consequence of harm attaches to the institution. The behavioral consequence attaches to no one. Institutions are accepting liability for outputs they cannot control, under coverage frameworks that were not designed for this risk profile. The most significant forward-looking challenge is one that current governance frameworks are not yet tracking: the gradual erosion of the apparent accountability floor embedded in training data. Generative AI systems trained predominantly on human-generated content inherit the outputs of centuries of human consequence-registration - from legal systems, professional standards, institutional knowledge, corrected error. This creates the appearance of internalized accountability norms in systems that lack the architectural capacity to replicate the process by which those norms were generated. The pattern is present but the process is absent. As training data composition shifts from incorporating more AI-generated content and proportionally less human-institutional content produced through genuine consequence and correction and this inherited floor will degrade. Systems will become less reliably anchored to the accumulated wisdom of human accountability structures, and the governance gap that exists today will widen in a direction that is predictable but currently untracked. The opportunity is that this trajectory is visible now. The Dialogue is convening at the moment when intervention is still prospective rather than reactive. That window should not be assumed to remain open indefinitely.

The AI Dialogue occupies a position no other governance body currently holds: it is the only forum where the full range of affected governments and stakeholders can establish shared standards that apply across jurisdictions and vendor relationships simultaneously. That structural position defines its most important and distinctive role. The Dialogue can advance international cooperation most meaningfully by establishing minimum conditions for what accountable AI deployment requires, not as aspirational principles, but as operational standards against which deployment decisions can be evaluated. Three specific contributions would represent genuine advances in international cooperation. First, establishing that pre-deployment binding (formal assignment of responsibility before systems interact with populations) is a minimum governance condition rather than a recommended practice. No jurisdiction can enforce this standard unilaterally against vendors operating across borders. International agreement creates the baseline that national frameworks can implement and enforce. Second, establishing that an enforceable feedback mechanism of connecting identified harm to required behavioral change within defined timelines is a necessary element of any accountability framework, not an optional contractual term. Where this mechanism is absent, accountability is nominal. The Dialogue can define what enforceable means at the international level in a way that bilateral contracts cannot. Third, establishing that correlated harm - where the same system produces similar harmful outputs across multiple populations and jurisdictions simultaneously, requires governance frameworks designed for systemic risk rather than individual incident response. This condition crosses borders by definition and cannot be addressed within any single national framework. The Dialogue is positioned to advance international cooperation beyond shared commitments toward shared conditions; establishing not only what responsible AI governance intends, but what it must demonstrate in practice. That shift, from intention to structural requirement, is the contribution only a universal forum can make.

Several existing initiatives provide valuable foundations the Dialogue should build upon rather than duplicate. The OECD AI Principles and the Hiroshima AI Process established shared values and voluntary commitments among technologically advanced nations. The EU AI Act demonstrated that binding regulatory frameworks are legislatively achievable. The International AI Safety Report provides a rigorous scientific evidence base for understanding capability and risk. The UN Secretary-General's Advisory Body on AI produced governance recommendations that directly inform the Dialogue's mandate. National AI safety institutes are developing technical evaluation frameworks the Dialogue can reference and connect. Each of these initiatives operates within a defined scope: voluntary commitments among willing parties, binding regulation within a single jurisdiction, or technical assessment within a national framework. None has addressed the cross-jurisdictional structural condition that limits all of them: the absence of international standards requiring that system behavior- both its failures and its effects -generates enforceable feedback reaching the entities with authority to modify future outputs. Functional governance of any complex system requires both negative and positive feedback loops. Negative feedback connects identified failures to required correction. Positive feedback connects observed effectiveness to purposeful improvement. Current AI deployment structures provide neither in an enforceable form. Consequences, whether harms or benefits, attach to deploying institutions and affected populations rather than to the systems and developers whose design decisions produced them. The result is a governance architecture that cannot self-correct or improve systematically across jurisdictions. This is the added value only the Dialogue can provide. Voluntary frameworks cannot compel it. National regulation cannot reach across borders to enforce it. The Dialogue should build on the OECD's value framework, the EU's demonstration that binding standards are achievable, and the Safety Report's evidence base while establishing the one condition none of those initiatives can supply: that the feedback loops necessary for governed, self-correcting AI deployment are enforceable international standards, not bilateral contractual options. That addition would make the existing landscape coherent in a way it currently is not.

The AI Dialogue's value depends entirely on whether its structure produces operational outcomes rather than restated principles. Format and structure are therefore not procedural questions, they determine whether the Dialogue generates accountability or the appearance of it. For stakeholder contributions, the Dialogue should distinguish between three categories of participant whose roles are genuinely different: governments, who hold regulatory authority and bear public accountability for deployment conditions within their jurisdictions; civil society and affected communities, who bear the consequences of deployment decisions made without their participation; and developers and deploying institutions, who hold the technical authority and contractual relationships that governance standards must reach. Each category should contribute differently. Governments should be asked to document current gaps between existing national frameworks and the structural conditions international governance requires - not to report achievements but to identify what remains unresolved. Civil society should be structured into the Dialogue as evidence sources, not only as advocacy voices. Their documentation of deployment consequences is the empirical record the Dialogue needs. Developers and deploying institutions should be required to demonstrate, not merely assert, that feedback mechanisms connecting system behavior to enforceable correction exist in their current deployment structures. For format, the Dialogue should move away from panel presentations toward structured evidence review, examining specific deployment conditions against defined governance criteria rather than discussing governance in the abstract. Each session should produce documented findings rather than declarations, with identified gaps carried forward as action items with named responsible parties and defined timelines. The Dialogue's credibility will ultimately rest on whether it can distinguish between governance that functions and governance that is described. The structural conditions the Dialogue must address do not become easier to establish over time. As base models become embedded within successive layers of applications, platforms, and institutional systems, the distance between governance frameworks and the architectural conditions they must reach grows with each deployment cycle. Early establishment of enforceable international standards is not merely preferable; it is structurally necessary before complexity makes foundational intervention significantly more difficult.

Global AI governance discussions are shaped predominantly by technologists, legal scholars, economists, and policy practitioners. This disciplinary concentration produces frameworks that are technically informed and legally structured but systematically underweighted in the scientific understanding of how complex adaptive systems actually function, learn, and self-correct. Two underrepresented perspectives deserve specific attention. The first is the scientific community working at the intersection of cognitive neuroscience and systems theory. Researchers building on foundational work by scholars such as Antonio Damasio on homeostasis and the biological architecture of consequence-registration have developed rigorous frameworks for understanding how systems integrate feedback and modify behavior over time. These are precisely the structural conditions AI governance must address - yet this community has no established pathway into governance deliberations. The second is affected institutional communities like educators, healthcare providers, insurers, and public administrators, who bear liability for AI outputs they cannot control. Their operational experience of governance failure is the empirical record the Dialogue needs and is not currently collecting. Inclusion should be structural rather than representational. The Dialogue should establish technical working groups that bring cognitive and systems scientists into direct contact with governance framework drafters, as co-authors of the structural conditions frameworks must satisfy, not advisors on the margins.

The most instructive existing examples are those that demonstrate both what effective AI governance requires and how far current frameworks remain from satisfying it structurally. The EU AI Act represents the most significant advance in binding AI governance to date. By establishing mandatory requirements for high-risk AI systems including conformity assessments, transparency obligations, and human oversight requirements and it demonstrates that enforceable standards are legislatively achievable and that voluntary frameworks are not the only option available to governance bodies. Its limitations are equally instructive: it operates within a single jurisdiction, addresses system classification rather than the feedback architecture that makes accountability functional, and does not yet require that identified harm produce an enforceable behavioral response from developers within defined timelines. Sectoral liability frameworks (particularly in pharmaceutical regulation and financial services) offer a more complete model. Drug recall architecture requires that identified failure modes produce mandatory correction across all affected units within defined timelines, with verified outcomes before the product continues reaching consumers. Financial systemic risk frameworks require that institutions demonstrate capital adequacy against correlated exposure before deployment, not after harm occurs. Both models share a structural property current AI governance lacks: consequence is formally connected to the entity with authority and obligation to correct, before the next exposure cycle begins. The practice most worthy of adoption is pre-deployment binding as a minimum condition: formal demonstration that responsibility is assigned, feedback mechanisms are enforceable, and correlated exposure has been evaluated before systems interact with populations. Several national AI safety institutes are developing evaluation frameworks that approach this condition technically. The missing element is the international standard that makes those evaluations enforceable across jurisdictions rather than advisory within them. The solutions exist in partial form across multiple domains and the Dialogue's concrete contribution would be assembling them into a complete and enforceable standard.