Graycon BV

Success cannot be measured by the production of a document. The history of multilateral governance is littered with well-worded principles that bound no one and changed nothing. The first Global Dialogue on AI Governance will be a success only if it produces architecture, not rhetoric. Three outcomes would constitute genuine success. First, agreement on a disaggregated problem structure. AI governance currently conflates safety risks, power concentration, and human rights into a single conversation, producing the lowest common denominator. A successful Dialogue establishes distinct tracks with distinct mandates, actor configurations, and accountability mechanisms. This is the precondition for everything else. Second, the institutionalisation of the consultative process itself. The Dialogue must not end with a summit. It must establish a permanent feedback mechanism, one that keeps governance current as the technology evolves, and that integrates civil society, sector actors, and crucially, Global South representatives as genuine design participants rather than implementation recipients. Buy-in is not built through consultation after decisions are made. It is built through participation in making them. Third, a credible pathway to a binding legal instrument. Not the instrument itself, as that is politically premature. But a clear, agreed roadmap toward a framework convention with national implementation obligations, modelled on existing multilateral treaty architecture. States that leave the Dialogue knowing what they are committing to, and when, represent a more valuable outcome than a consensus communiqué that commits no one to anything. What would make the Dialogue a failure is equally clear: a declaration of principles with no enforcement pathway, a process that excludes the Global South from design authority, and a framework that addresses civilian AI while leaving the military dimension unresolved and therefore ungovernable. The window is open. The question is whether the Dialogue uses it.

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

These four areas are not independent priorities. They form an interconnected architecture, and that interconnection is precisely why they require urgent action together rather than sequentially. Safe, secure and trustworthy AI is the entry point, but safety cannot be defined in the abstract. It must be defined in relation to who is protected, under what conditions, and by whom. That requires the human rights framework and the ethical and social implications lens to be present from the start, not added later as a correction. The social, economic, ethical, cultural, and linguistic implications of AI are where governance either becomes real or remains theoretical. This is where gender inequalities are encoded or prevented, where Global South communities experience AI as something done to them rather than designed with them, and where dual-use technologies cause harm that civilian governance frameworks currently cannot reach. Protection and promotion of human rights provides the normative floor. Without it, safety and trustworthiness become industry-defined concepts subject to commercial interests. Transparency, accountability, and human oversight are the enforcement mechanisms that make the other three meaningful. Principles without accountability are declarations. Accountability without transparency cannot be verified. Human oversight is the check that prevents both state and corporate actors from operating AI systems beyond democratic scrutiny. These four areas together constitute the minimum viable architecture for governance that is legitimate, enforceable, and structurally resistant to capture.

Two cross-cutting issues are insufficiently captured by the listed themes and require explicit integration into the Global Dialogue. The first is the dual-use nature of AI, particularly its military dimension. AI is not a civilian technology with occasional military applications. It is an inherently dual-use technology in which the same systems, models, and infrastructure underpin both domains simultaneously. A governance framework that addresses only civilian applications contains a structural vulnerability: actors seeking to avoid regulatory obligations can reclassify applications as falling within the defence domain. This renders civilian governance permanently porous. The military dimension is therefore not a parallel conversation to be handled separately. It is a precondition for the integrity of civilian governance. The work of the Group of Governmental Experts on Lethal Autonomous Weapons Systems must be formally connected to this Dialogue, not left running as a disconnected technical process. Export controls on AI-capable technologies must also be explicitly addressed. The second is the question of who produces knowledge about AI governance. The listed themes implicitly assume that governance expertise originates in the Global North and is subsequently applied elsewhere. This assumption is both empirically incorrect and politically consequential. Communities in the Global South have direct, granular experience of AI deployment in contexts, including humanitarian operations, agricultural systems, and public health infrastructure, that generate knowledge unavailable to regulators in Brussels or Washington. That knowledge must be formally recognised as a legitimate input to governance design, not treated as a regional implementation concern. Both issues share a common structure: they are not thematic gaps so much as architectural blind spots. Leaving them unaddressed does not simply make the framework incomplete. It makes it exploitable.

The European Union occupies a structurally distinctive position in global AI governance. It is neither a primary AI producer nor a passive recipient of AI systems developed elsewhere. It is a regulatory actor with significant normative reach, demonstrated by the extraterritorial effect of the General Data Protection Regulation and, more recently, the AI Act. This position generates both opportunity and a specific set of governance vulnerabilities. The most significant challenge is the gap between regulatory ambition and enforcement capacity. The AI Act establishes a sophisticated risk-based framework, but its effectiveness depends on conformity assessment infrastructure, market surveillance capacity, and cross-border coordination that does not yet exist at the required scale. Governance instruments are being written faster than the institutional capacity to apply them is being built. The second challenge is fragmentation. The EU framework, the Council of Europe AI Convention, bilateral agreements, and sector-specific instruments are developing in parallel without adequate interoperability. For practitioners operating across these frameworks, the result is legal uncertainty that paradoxically advantages large actors with the compliance resources to navigate complexity, and disadvantages smaller organisations and civil society. The opportunity is equally significant. Europe has demonstrated that a sufficiently large regulatory coalition can generate normative gravity that extends well beyond its borders. The GDPR became a de facto global standard not through universality but through market access conditions. The same mechanism is available for AI governance if the EU coordinates its external engagement with the Global Dialogue rather than treating them as separate tracks. For civil society actors in particular, the current moment offers genuine influence. The governance architecture is not yet fixed. The decisions made in this Dialogue will shape the institutional landscape for decades. That window will not remain open.

The AI Dialogue occupies a position in the international governance landscape that no existing institution currently fills. It is neither a treaty body nor a technical standards organisation. That ambiguity is not a weakness. It is, if used deliberately, its most significant asset. Existing AI governance efforts are fragmented by design. The EU AI Act governs one jurisdiction. The Council of Europe Convention is regional. The G7 Hiroshima Process is club-based. The GGE on Lethal Autonomous Weapons operates in a separate silo. Each addresses a portion of the problem. None has the legitimacy or the mandate to connect them. The Dialogue can play three distinct roles in advancing international cooperation that no other forum can currently claim. First, it can function as a mapping and convergence mechanism, identifying where existing frameworks are compatible and where they conflict, and building the normative common ground necessary for a future binding instrument. This is not a modest ambition. The absence of that map is itself a governance gap that industry actors exploit. Second, it can serve as the primary legitimation channel for Global South participation in governance design. No other current process has the universality of the UN system combined with a mandate to engage non-state actors. That combination is rare and should not be squandered on producing declarations. Third, it can establish the consultative infrastructure that makes governance adaptive rather than reactive. If the Dialogue institutionalises a permanent feedback mechanism connecting field experience, civil society knowledge, and technical developments to the bodies that write governance instruments, it will have built something more durable than any single agreement. The Dialogue's value is not what it decides. It is what it makes possible for the processes that follow.

Several existing initiatives provide foundations the Dialogue should explicitly build upon rather than duplicate. The EU AI Act represents the most comprehensive binding regulatory framework currently in force. Its risk-based architecture, conformity assessment mechanisms, and fundamental rights impact requirements offer a tested model for translating governance principles into enforceable obligations. The Dialogue should treat it as a reference point for treaty design, not as a regional instrument irrelevant to global governance. The Council of Europe Framework Convention on AI, is the first binding international instrument on AI and human rights. Its membership extends beyond Europe to non-European signatories. The Dialogue must connect to this process explicitly, or risk producing a parallel instrument that fragments rather than consolidates the normative landscape. The GGE on Lethal Autonomous Weapons Systems under the Convention on Certain Conventional Weapons has been addressing the military dimension of autonomous systems for over a decade. Its work must be formally integrated into the Dialogue architecture. Continued parallel operation produces exploitable incoherence between civilian and military governance tracks. The OECD AI Principles and the G7 Hiroshima Process provide existing multilateral consensus among major AI-producing states. They are non-binding, but they represent negotiated common ground that the Dialogue can build upon toward harder instruments. The UNESCO Recommendation on the Ethics of AI, provides the broadest existing normative consensus and the most explicit integration of gender and cultural diversity dimensions. The added value the Dialogue brings to all of these is singular: universality combined with a multi-stakeholder mandate. No existing initiative connects all of these frameworks, all UN member states, and civil society within a single process. That connective function, if institutionalised rather than performed episodically, is what transforms the Dialogue from a conversation into infrastructure.

The format and structure of the AI Dialogue will determine whether its outputs are legitimate. Process is not separate from substance here. Who participates, in what capacity, and at what stage of decision-making are governance questions as consequential as the content of any agreement produced. Four structural recommendations follow. First, participation must be organised into defined clusters rather than open-ended stakeholder categories. Civil society, industry, academia, technical standards bodies, and government actors each bring distinct knowledge and distinct interests. Conflating them into a single stakeholder category obscures those differences and allows better-resourced actors to dominate. Cluster-based participation with transparent selection criteria and balanced geographic representation is the minimum viable architecture for legitimate multi-stakeholder engagement. Second, Global South actors must participate as co-designers, not consultees. This requires structural measures beyond invitation: dedicated preparation support, language accessibility across all working documents, and formal mechanisms that give Global South contributions equal weight in drafting processes. Participation without influence is performance. Third, industry actors should participate within a clearly defined and separate track, with explicit disclosure obligations regarding commercial interests. Their technical knowledge is necessary. Their role as subjects of governance, rather than architects of it, must be structurally enforced rather than aspirationally stated. Fourth, the Dialogue must establish a permanent secretariat with a standing consultative mechanism, not a conference cycle. Periodic summits produce periodic outputs. A standing mechanism with regular structured input from all stakeholder clusters, feeding directly into governance review processes, produces adaptive governance. The AI for Good Summit in Geneva provides an existing annual infrastructure that could anchor this function if given an explicit governance mandate rather than remaining primarily a showcase event. The Dialogue's legitimacy will be judged not by who was invited, but by who was heard, and what changed as a result.

The underrepresentation in global AI governance discussions is not incidental. It reflects the same structural inequalities that AI systems themselves risk encoding. Addressing it requires more than expanded invitation lists. Six groups are systematically underrepresented, each for distinct structural reasons. Communities in the Global South are present primarily as case studies of harm rather than as architects of governing frameworks. Closing this gap requires shifting design authority, not just participation rights, to include actors with direct field experience of AI deployment in contexts that Northern regulators do not encounter. Women and gender-diverse communities are disproportionately affected by AI across domains including labour displacement, targeted harassment, and biometric surveillance, yet remain underrepresented in technical and policy-making bodies. Mandatory gender balance and gender-disaggregated impact assessment must be structural requirements, not aspirational targets. LGBTQ+ communities face specific risks from AI systems that infer sexual orientation or gender identity, suppress LGBTQ+ expression through content moderation, or enable state surveillance targeting. Governance frameworks must address these risks explicitly and create protected channels for input from communities that cannot safely participate under their own names. People with disabilities are both disproportionately affected by AI deployment in healthcare, employment screening, and public services, and disproportionately absent from governance design. Accessibility must be treated as a design requirement for participation processes, not an accommodation added afterwards. Indigenous communities hold knowledge about data sovereignty and collective rights directly relevant to AI governance. Their participation requires recognition of collective rather than only individual rights frameworks. Workers in the informal economy and in AI-adjacent labour such as content moderation experience governance failures most immediately and remain almost entirely absent from formal processes. Inclusion means structured access, preparation support, and formal influence over outcomes. Anything less is consultation without consequence.

The risk of the AI Dialogue is not insufficient ambition. It is insufficient imagination about how multilateral processes actually produce knowledge and commitment. Standard formats, keynote addresses, panel discussions, written submissions, systematically advantage actors who are already fluent in multilateral procedure and disadvantage precisely the communities whose participation the Dialogue most needs. Four innovative engagement formats would meaningfully address this. Structured deliberative panels drawn from underrepresented communities, with professional facilitation and guaranteed input into formal outputs, would replace the current model in which civil society speaks into a process and then waits to see whether it was heard. Deliberative formats are not new. Applying them systematically within a UN process at this scale would be. Red team sessions in which civil society actors, technologists, and affected communities are tasked with identifying exploitable gaps in proposed governance frameworks would produce more robust instruments than consultation rounds. Adversarial scrutiny by design is more effective than goodwill by assumption. Asynchronous and multilingual input mechanisms that allow substantive contributions outside the time zones, languages, and institutional formats of formal sessions would expand the geographic and demographic range of participation without requiring physical or synchronous presence. Written submissions are already used but remain inaccessible to many communities without dedicated support. Living document review processes in which draft governance texts are opened to structured iterative comment from defined stakeholder clusters, rather than presented as finished products for endorsement, would shift the Dialogue from a validation exercise to a genuine co-design process. Underlying all of these is a single principle: engagement formats must be designed around the participants the Dialogue most needs to reach, not around the convenience of those already at the table. Format is not a logistical question. It is a governance question.

Several existing instruments demonstrate that effective AI governance is achievable when it combines binding obligations, adaptive mechanisms, and genuine multi-stakeholder input. The EU AI Act is the most comprehensive binding regulatory framework currently in operation. Its risk-based architecture, which calibrates obligations to the severity of potential harm, offers a replicable model for translating principles into enforceable requirements. Its mandatory fundamental rights impact assessments for high-risk systems demonstrate that human rights protection can be built into compliance architecture rather than added as an afterthought. The GDPR demonstrates the mechanism of normative gravity: a jurisdiction-specific instrument that became a de facto global standard through market access conditions rather than universal ratification. This is the most important proof of concept available to the Global Dialogue for how binding governance can achieve global reach without requiring simultaneous universal agreement. The MARPOL Convention for maritime pollution demonstrates that a framework convention can accommodate rapid technical evolution through structured annex procedures, allowing specific standards to be updated without renegotiating the core treaty. This model is directly applicable to AI governance, where technical development will consistently outpace conventional treaty cycles. The Aarhus Convention on environmental governance offers a model for institutionalised civil society participation with genuine legal standing, including the right to challenge governance decisions. Applied to AI, this would mean civil society actors having formal recourse when governance obligations are not met, rather than relying on political pressure alone. Finally, the Open Government Partnership demonstrates that voluntary commitments combined with peer review and civil society monitoring can generate meaningful compliance incentives even without hard enforcement mechanisms. For states not yet ready for binding commitments, this model offers a pathway toward accountability. Together these examples share a common feature: they treat governance as infrastructure to be maintained and updated, not as a problem to be solved once and archived.