information labs

Success requires one thing above all: a Dialogue that treats AI governance as a structural problem rather than a coordination problem. Most existing international AI processes have failed not because governments disagreed on values, but because the forums were designed to exchange positions rather than resolve conflicts between them. A first session that produces shared principles will have achieved very little. A first session that produces a map of where governance approaches are genuinely incompatible, and a method for resolving that incompatibility, will have achieved something. Three concrete markers of success: First, the Dialogue should establish a baseline definition of what "interoperable" means when applied to AI governance frameworks. Interoperability between the EU AI Act, the US executive order landscape, China's algorithmic regulation, and the African Union's framework is not self-evident. Without a shared definition, the word functions as aspiration rather than architecture. Second, the process should produce a capacity gap audit: a frank account of which governments currently lack the technical infrastructure, legal expertise, and institutional capacity to implement any of the thematic priorities. Capacity-building commitments made without this audit will misallocate resources. Third, civil society and the technical community should leave Geneva with standing commitments from member states on how their inputs will be reflected in outcomes. The submission process generates a record. The Dialogue should generate accountability mechanisms tied to that record. What the Dialogue should explicitly avoid: producing a declaration that rehearses the language of the Global Digital Compact without advancing beyond it. The Pact for the Future established the principle that AI governance is a shared international responsibility. The Dialogue's job must be to operationalise that principle, not to merely reaffirm it.

AI capacity-building;Transparency, accountability, and human oversight;Interoperability of governance approaches;Social, economic, ethical, cultural, linguistic and technical implications of AI;

These four themes are selected because they address the execution gap in AI governance rather than the aspiration gap. Governments and international bodies have produced substantial agreement on AI values. The gap is between stated values and governing instruments capable of implementing them. Transparency, accountability, and human oversight sit at the centre of every credible AI governance framework, yet remain the least operationalised. The EU AI Act defines conformity requirements for high-risk systems, but auditing mechanisms remain immature, and third-party audit capacity does not exist at scale. The Dialogue should prioritise shared audit standards over shared principles. AI capacity-building is the precondition for meaningful participation in every other theme. 118 countries are currently absent from prominent international AI governance initiatives. Their absence is not a product of indifference. It reflects the absence of the institutional infrastructure required to engage. Capacity-building that focuses on regulatory bodies and technical agencies within those governments, rather than on generic digital literacy, will produce durable results. Interoperability of governance approaches is the least developed theme on the list and the one where international process adds the most value. National and regional frameworks are being built at speed, and the compatibility decisions being made now will determine whether global AI supply chains face regulatory fragmentation equivalent to the data localisation conflicts of the 2010s. The Dialogue should focus on layer-by-layer interoperability: where frameworks can be made compatible at the definition level, the audit level, and the enforcement level respectively. Social, economic, ethical, cultural, and linguistic implications round out the four because AI's distributional effects are already visible and already unequal. Governance that attends only to safety and security, without attending to who captures AI's benefits and who bears its costs, is not neutral. It is a policy choice in favour of incumbents.

Two issues warrant explicit inclusion. The first is AI's relationship to information integrity and democratic infrastructure. The listed themes address what AI systems do and how they are built. None of them directly addresses what AI systems do to the epistemic conditions that make governance possible. Generative AI is already reshaping the information environments in which elections are conducted, in which public health guidance is received, and in which policy debates occur. The effects are not hypothetical: AI-generated synthetic media, micro-targeted political messaging, and algorithmically amplified disinformation are active and documented. A Dialogue on AI governance that does not address AI's impact on the conditions for informed consent and democratic deliberation has a significant blind spot. The second is the concentration of AI infrastructure. The Dialogue's thematic areas assume a landscape of national and regional regulatory actors governing AI systems developed across a distributed ecosystem. The actual landscape is substantially different. A small number of corporations control the compute infrastructure, the foundational models, and the data pipelines on which virtually all downstream AI applications depend. Regulatory frameworks that operate at the application layer while leaving the infrastructure layer ungoverned will systematically underperform. International governance processes have confronted this structure before, in telecommunications and in internet infrastructure, and have produced mixed results. The Dialogue should name the concentration problem explicitly and assess whether existing thematic priorities are sufficient to address it. These two issues are connected. Concentrated infrastructure and degraded information environments are not separate governance problems. They share a causal structure: systems optimised for engagement and scaled through concentrated infrastructure produce information environments that undermine the public reasoning capacity on which all other governance depends. A Dialogue that addresses AI's technical architecture without addressing its epistemic consequences will produce incomplete governance.

The EU is the jurisdiction that has moved furthest and fastest on AI governance, and that position generates a specific set of challenges that the Dialogue should understand rather than simply celebrate. The EU AI Act is the most comprehensive binding AI regulatory framework in existence. Its risk-based architecture, conformity requirements for high-risk systems, and prohibitions on unacceptable-risk applications represent a genuine regulatory achievement. The implementation reality is considerably more complicated. The Act's obligations are landing on member state market surveillance authorities that lack the technical capacity to enforce them, on small and medium enterprises that lack the legal capacity to comply with them, and on a conformity assessment infrastructure that does not yet exist at the scale the regulation requires. A framework that is architecturally sound but practically unenforceable produces the worst of both outcomes: regulatory burden without regulatory protection. The second challenge is jurisdictional asymmetry. The EU regulates AI systems deployed within its borders, but the foundational models, training infrastructure, and data pipelines underlying those systems are predominantly developed and controlled outside them. Extraterritorial reach through the Act's market access provisions addresses part of this gap. It does not address the deeper question of whether application-layer regulation can govern systems whose consequential decisions are made at the infrastructure layer. The opportunity the EU brings to the Dialogue is a working legal text, an enforcement architecture in construction, and three years of implementation experience that other jurisdictions do not yet have. That experience, including its failures, is more useful to international governance than the EU's published principles. The Dialogue should treat the EU not as a model to be replicated but as a field experiment whose results, positive and negative, should be systematically shared with governments building frameworks from a different starting position.

The Dialogue's most valuable and least replicable function is to serve as the venue where governance frameworks that were built independently are made to confront each other. The G7, G20, OECD, Council of Europe, African Union, and ASEAN have all produced AI governance instruments. None of those processes had universal membership. None was designed to resolve conflicts between frameworks developed elsewhere. The result is a growing body of governance activity that is internally coherent and externally incompatible. The Dialogue is the only forum with both the membership and the mandate to address that incompatibility directly. It should not use that position to produce a lowest-common-denominator synthesis. It should use it to map, with precision, where frameworks diverge at the definition level, the audit level, and the enforcement level, and to establish processes for resolving those divergences in ways that do not simply defer to the most powerful regulatory bloc. A second and equally important role is asymmetric knowledge transfer, running in both directions. The dominant assumption in international AI governance is that knowledge flows from technically advanced jurisdictions to developing ones. That assumption is wrong in at least two respects. Jurisdictions that have governed technology deployment under conditions of limited infrastructure have developed regulatory approaches, particularly around mobile-first AI applications and community-level impact assessment, that technically advanced jurisdictions have not needed to develop and do not have. The Dialogue should be designed to surface that knowledge systematically, not as a gesture toward inclusion but because the governance problems AI will generate at scale in the next decade will look more like the problems already being managed in the Global South than like the problems currently preoccupying Brussels and Washington.

The landscape of existing AI governance initiatives is extensive and underconnected. The Dialogue should build on four in particular, for distinct reasons. The OECD AI Policy Observatory has produced the most comprehensive comparative database of national AI strategies, regulatory frameworks, and policy developments currently in existence. The Dialogue should treat this as its evidentiary baseline rather than commissioning parallel work. The added value the Dialogue brings is universality: the Observatory's membership skews toward high-income countries, and the Dialogue can extend its analytical framework to jurisdictions the Observatory does not adequately cover. The Council of Europe Framework Convention on AI is the only binding international legal instrument on AI and human rights currently in force. The Dialogue is not a treaty-making body, but it should actively engage with the Convention's implementation record. Jurisdictions outside the Council of Europe are watching whether binding obligations produce better outcomes than voluntary principles. That question will shape the appetite for harder governance instruments globally, and the Dialogue is the right venue to assess the evidence. The Global Partnership on AI brought together fourteen founding members in 2020 with an explicit mandate to bridge AI development and governance. Its absorption into the OECD structure reduced its profile without resolving its mandate. The Dialogue should recover that mandate and give it universal standing. The AI Safety Institutes network, now spanning more than a dozen national bodies, represents the most technically capable governance infrastructure currently operating. Its focus on frontier model risk is narrower than the Dialogue's thematic scope, but its methods, particularly its approach to pre-deployment evaluation, are transferable to a wider range of governance questions. The Dialogue should formalise the relationship between national safety institutes and the international process rather than allowing them to develop as parallel tracks.

The Dialogue's format should be determined by its function, and its function is to resolve conflicts between governance frameworks, not to showcase them. That requires a structure different from the standard international forum model, in which governments present positions, civil society observes, and industry hosts side events. Three structural recommendations follow from that premise. Governments should enter each session with documented positions on specific interoperability questions, not general statements of principle. The pre-session submission process should require member states to identify, by name, at least one governance conflict with another jurisdiction's framework and their preferred resolution. This converts the Dialogue from a venue for position-stating into a venue for negotiation. Civil society and independent research organisations should have standing to submit evidence, not merely to observe proceedings. The distinction matters. Observer status produces testimony. Evidence standing produces a record that session outcomes must address. The EU Digital Services Act's vetted researcher framework offers a workable model for how independent researchers can be given structured access to proceedings and outputs without compromising intergovernmental negotiating space. Industry participants should be required to disclose, as a condition of participation, which national regulatory frameworks their systems are currently subject to and the status of their compliance with each. Self-disclosure of this kind, published before each session, would give other participants the factual basis for assessing industry positions on interoperability and regulatory burden. The two-day annual format established by Resolution 79/325 is insufficient for the work the Dialogue is being asked to do. Intersessional working groups, structured around specific thematic conflicts rather than the full agenda, should operate continuously between sessions and report outcomes rather than progress updates. A forum that meets for two days a year and produces declarations will not keep pace with the technology it is governing.

Five groups are systematically underrepresented in current global AI governance discussions, for different reasons and requiring different remedies. Women are underrepresented at every layer of the AI economy: in the research and engineering roles that shape system design, in the regulatory and policy roles that shape governance, and in the forums where international standards are set. The consequences are not merely procedural. AI systems trained on datasets that underrepresent women, evaluated by teams that underrepresent women, and governed by processes that underrepresent women will systematically encode and amplify that absence. Work currently underway through initiatives such as Connecting Women in Digital's Working Group on Women in the AI Economy documents both the scale of the participation gap and the specific barriers, including unpaid care obligations, pipeline exclusion, and the concentration of AI investment in male-dominated networks, that sustain it. That evidence base should inform the Dialogue's design, not sit outside it. Governments of low-income countries are formally present in UN processes but substantively marginal. Dedicated technical support for delegations, funded through the Dialogue's core budget rather than voluntary contributions from interested parties, is the minimum condition for genuine engagement. Voluntary funding from industry or technically advanced governments creates conflicts of interest that constrain the positions those delegations can credibly take. Affected communities, meaning people whose employment, access to public services, and legal standing are directly shaped by AI deployment, are represented at best through civil society proxies. Structured participatory mechanisms, drawing on deliberative democracy methods already used in technology governance in Ireland, France, and Canada, should be built into the Dialogue's design from the outset. Workers in AI-intensive sectors in the Global South, particularly content moderators and data labellers, have direct material stakes in governance outcomes and negligible presence in the forums that determine them. Labour organisations representing these workers should hold formal standing, not side-event access.

The most productive innovation the Dialogue could adopt is also the least glamorous: structured disagreement as a formal agenda item. Current international AI governance forums are organised around consensus-building. Sessions are designed to surface areas of agreement and defer areas of conflict. This produces documents that are broadly endorsed and operationally inert. The Dialogue should invert this design. Each session should include formal agenda time dedicated to mapping specific, named governance conflicts, presenting the strongest version of each position, and producing a public record of where disagreement lies and why. This is not a failure mode. It is the precondition for the negotiation that produces durable governance. Red team panels, in which a designated group is tasked with identifying the weakest points of proposed governance frameworks, should be a standing feature of each session. The panels should include technical experts, affected community representatives, and legal scholars, and their findings should be published alongside, not instead of, the frameworks they assess. Governance processes that are not stress-tested before implementation are stress-tested by events. Between sessions, the Dialogue should pilot structured bilateral exchanges between regulatory bodies in different jurisdictions working on equivalent governance problems. The exchange between the EU's AI Office and equivalent bodies in the African Union, India, and Brazil would produce more actionable knowledge than an additional plenary session. These exchanges should be documented and published as part of the Dialogue's record. Finally, the Dialogue should commission independent retrospective assessments of its own sessions, conducted by researchers with no involvement in the process, and publish those assessments before the following session. International governance forums that evaluate their own effectiveness tend to find that they are effective. Assessments conducted at arm's length tend to find otherwise. The credibility of the Dialogue depends on which model it chooses.

Four examples stand out, each for a different reason. The EU AI Act's risk-based architecture is the most architecturally coherent binding framework currently in force. Its contribution to international governance is not the specific obligations it creates but the analytical method it establishes: classifying systems by the severity and reversibility of harm they can cause, and calibrating regulatory requirements accordingly. That method is transferable to jurisdictions with very different legal traditions and institutional capacities. What is not transferable without adaptation is the conformity assessment infrastructure the Act assumes, which does not exist at scale even within the EU. The lesson for other jurisdictions is to adopt the risk classification logic while building the audit infrastructure before, not after, the regulatory obligations land. The EU Digital Services Act's vetted researcher access framework is the most promising model currently in operation for independent oversight of platform systems. Allowing researchers with verified independence to access data and algorithmic systems that would otherwise remain opaque addresses the evidentiary gap that has hampered AI governance research for a decade. Its limitations, including slow implementation and platform resistance, are instructive. Access rights without enforcement mechanisms are aspirational. Rwanda's national AI policy, published in 2023, offers a model for how a government with limited technical infrastructure can produce a governance framework that is realistic about its own constraints, focused on specific high-value applications, and designed around the institutional capacity that actually exists rather than the capacity that international standards assume. It is the most honest national AI strategy currently in circulation. Chile's algorithmic impact assessment requirement for public sector AI systems represents the clearest example of transparency obligations being operationalised at the procurement stage rather than retrofitted after deployment. Embedding governance requirements into public procurement is faster, cheaper, and more reliably enforced than post-deployment audit regimes. It deserves wider adoption.