Microsoft

Success would mean the Dialogue produces something more than a communiqué that reaffirms existing commitments. For this to be a genuinely transformative moment, three things need to happen. The Dialogue must surface governance realities that are currently absent from multilateral AI discussions. Much of what passes for "global" AI governance is, in practice, a negotiation between a small number of technologically advanced states and large platform companies. The experiences of countries like Kenya, Rwanda, and Uganda — which are actively deploying AI in public services, urban infrastructure, and financial systems — are rarely integrated into frameworks at the design stage. A successful outcome would establish a mechanism for these contexts to inform, not merely receive, global governance norms. The Dialogue should also produce actionable guidance on governance interoperability. Different regulatory approaches are emerging across regions and the risk of fragmentation is real. Success here is not harmonisation for its own sake, but a shared methodology for identifying where approaches can coexist and where divergence creates harm for the communities most exposed to ungoverned AI systems. Most importantly, the Dialogue should be honest about who is missing. Civil society organisations, researchers working outside major institutions, and community-level actors rarely make it into rooms like Geneva. A successful Dialogue would identify concrete steps to change that — not as aspiration but as structured commitment with named accountability. If participants leave Geneva with a stronger and more honest understanding of what inclusive AI governance actually requires across different political economies and infrastructure realities, that would represent genuine progress worth building on.

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

These four areas directly reflect the findings of multi-country research I have conducted examining how AI governance functions and frequently breaks down in practice across Nairobi, Kigali, Lagos, and Cape Town. The social, economic, ethical, cultural, and technical implications theme is the most urgent because governance frameworks are routinely designed without adequate understanding of the social contexts into which AI systems are deployed. In the cities I have studied, AI is being used to allocate public services, manage urban mobility, and inform judicial and financial decisions, often with minimal public deliberation and significant cultural mismatch between system assumptions and the lived realities of the people affected. Transparency and accountability matters because across these contexts the gap between formal accountability commitments and actual practice is wide and growing. Procurement processes obscure algorithmic decision-making, audit mechanisms are underdeveloped, and affected communities have limited recourse when things go wrong. Governance frameworks that do not address this operational gap remain largely symbolic regardless of how well intentioned they are. Interoperability of governance approaches is critical because the countries I have studied are simultaneously navigating national AI strategies, regional frameworks including the African Union's continental AI strategy, and significant pressure to align with EU and US-led approaches. This creates substantial compliance burden and regulatory confusion, particularly for smaller states with limited technical capacity to absorb competing frameworks. AI capacity-building underpins all of the above. Without sustained investment in local technical expertise, legal capacity, and civil society capability to engage meaningfully in governance processes, frameworks will remain dependent on external expertise that does not reflect local priorities or conditions.

One issue that cuts across every thematic area but is not adequately named anywhere in the current framework is the distribution of governance labour itself - specifically, who is doing the work of AI governance and under what conditions. Research I have conducted, including interviews with over forty practitioners across government, civil society, and the private sector in multiple countries, consistently finds that AI governance is not primarily enacted through legislation or formal regulatory bodies. It is carried out by a distributed and largely invisible workforce: compliance officers, ethics reviewers, data annotators, community liaison staff, and mid-level civil servants who interpret, adapt, and operationalise AI policies in their day-to-day work. This labour is systematically undervalued, chronically under-resourced, and almost entirely absent from global governance conversations despite being the practical mechanism through which governance either succeeds or fails. This matters for the Dialogue because frameworks that focus exclusively on high-level principles and institutional design will continue to underestimate what effective implementation actually requires on the ground. The gap between policy and practice in AI governance is not primarily a technical problem. It is a resource, capacity, and visibility problem that no amount of principle-setting will resolve if it remains unaddressed. A related issue is the cultural and epistemic assumptions embedded in dominant AI governance frameworks. Concepts like transparency, accountability, and human oversight carry different institutional meanings across legal traditions and cultural contexts. In settings where community-based dispute resolution, oral governance traditions, or non-Western conceptions of collective rights shape how accountability is understood, direct translation of OECD-derived frameworks produces systems that are formally compliant but substantively misaligned with the communities they are meant to serve. The Dialogue should consider how to build governance frameworks that are not merely translated across contexts but genuinely co-produced with the communities most affected by AI systems. That is what inclusive governance actually means in practice.

Across Sub-Saharan Africa, the governance gaps in the thematic areas I identified are producing a specific and underappreciated dynamic: AI systems are being deployed at scale in high-stakes domains while the governance infrastructure needed to make them accountable is still being assembled, often by the same institutions responsible for deployment. This is not a future risk. It is the present condition in which millions of people are already living. In the cities covered by my research — Nairobi, Kigali, Lagos, and Cape Town — AI systems are actively shaping access to credit, social protection, urban services, and in some cases policing. The communities most exposed to these systems are also the least represented in the governance processes designed to oversee them. This asymmetry is the defining governance challenge of the moment and it is not receiving the attention it deserves in multilateral spaces. The interoperability gap creates particular strain. African governments are simultaneously being asked to align with the EU AI Act, engage with AU continental frameworks, and maintain bilateral technology relationships with partners whose governance expectations differ significantly. For smaller states with limited legal and technical capacity, this is not a manageable situation. It produces regulatory paralysis or, more commonly, de facto deference to whichever external framework arrives with the most investment attached to it. The opportunity, which is real and should not be lost, is that several African governments are actively developing AI governance approaches that are contextually grounded and empirically informed. Rwanda's digital governance architecture and Kenya's emerging AI policy process both represent serious attempts to build accountability mechanisms that reflect local institutional realities. The Dialogue should find ways to amplify and learn from these efforts rather than position them as recipients of best practice from elsewhere.

The most valuable role the AI Dialogue can play is one that no existing mechanism is currently filling: creating a structured space where governance experiences from across the full range of political economies — not just the most technologically advanced — are treated as evidence rather than context. International cooperation on AI governance has so far been shaped primarily by a small number of actors with the institutional capacity and geopolitical weight to set agendas. The result is a body of emerging norms that reflects the priorities and assumptions of those actors more than it reflects the diversity of contexts in which AI is actually being deployed and experienced. The Dialogue can change this, but only if it is deliberately designed to do so rather than assuming that open participation produces genuine inclusion. Concretely, the Dialogue can advance cooperation by serving three functions that are currently fragmented across different institutions. It can act as a knowledge exchange mechanism that surfaces governance approaches developed outside OECD contexts and makes them available to the broader international community on equal terms. It can act as an early warning system that identifies where governance gaps are producing measurable harm before those harms become crises requiring emergency response. And it can act as a coordination forum that helps governments navigate the growing complexity of overlapping and sometimes contradictory governance frameworks without simply defaulting to the framework backed by the largest economy in the room. None of this requires the Dialogue to become a regulatory body or to produce binding outcomes. It requires it to be honest about the limits of existing cooperation efforts and serious about the structural changes needed to make future efforts more legitimate and more effective.

Several existing initiatives have done important groundwork that the Dialogue should engage with seriously rather than duplicate. The OECD AI Policy Observatory has developed comparative frameworks and indicator sets that provide a useful baseline for understanding governance approaches across member and partner countries. The African Union's Continental AI Strategy represents a substantive attempt to articulate governance priorities from an African perspective and deserves far more prominence in multilateral discussions than it currently receives. UNESCO's Recommendation on the Ethics of AI, adopted in 2021, established a normative foundation that the Dialogue should explicitly build on rather than treat as a parallel track. The Global Partnership on AI, despite its limitations in terms of geographic representation, has produced technical and policy research that remains relevant to several of the Dialogue's thematic areas. What the Dialogue can add that none of these mechanisms currently provides is a genuinely multi-stakeholder forum with the convening authority of the UN system behind it. The OECD's reach is limited by membership. The AU strategy is under-resourced and under-connected to global processes. UNESCO's recommendation lacks an implementation and accountability mechanism. GPAI has struggled to maintain momentum and representational breadth simultaneously. The added value of the Dialogue is therefore not another layer of principle-setting. It is the potential to connect these existing efforts into something more coherent, with particular attention to the actors currently sitting outside all of these processes. Researchers and civil society organisations working in contexts that none of the existing initiatives adequately represent have produced knowledge that is directly relevant to building governance frameworks that actually work. The Dialogue should create a formal pathway for that knowledge to enter and influence the international conversation rather than circulating only in academic journals and regional policy spaces.

The question of how different stakeholders can contribute is inseparable from the question of whether the Dialogue is genuinely designed to receive their contributions or merely to appear inclusive while remaining structured around the preferences of the actors who have historically dominated these spaces.Governments bring regulatory authority and the political legitimacy to translate dialogue outcomes into enforceable frameworks. Their participation is necessary but should not be sufficient to determine the agenda. The Dialogue should ensure that government delegations include technical and legal practitioners with implementation experience, not only senior officials whose relationship with AI governance is primarily rhetorical.Academic researchers and civil society organisations bring empirical grounding that intergovernmental processes consistently lack. Their contribution should be structured into the Dialogue at the design stage, not added as a consultation layer after the substantive agenda has already been set. This means giving researchers and civil society meaningful input into which questions the Dialogue is asking, not just inviting them to respond to questions framed elsewhere.The private sector brings technical knowledge and operational scale that no other stakeholder group can replicate. It also brings conflicts of interest that need to be named honestly. The Dialogue should create mechanisms for private sector input that are transparent about the interests involved and structured to prevent regulatory capture of the process.Most importantly, the Dialogue should create formal roles for community-level actors and practitioners who work directly with AI-affected populations. These are the people who understand what governance failures actually look like when they arrive in someone's life. Their knowledge is not anecdotal. It is the most direct evidence available about whether AI governance is working, and it should be treated accordingly.

The most significant representational gap in global AI governance discussions is not demographic, though demographic gaps are real and serious. It is epistemic. The frameworks, concepts, and analytical categories that dominate global AI governance have been produced primarily within a small set of institutional contexts — Northern American and European universities, large technology companies, and OECD-aligned policy bodies. Other ways of understanding accountability, collective rights, harm, and trust are present in governance conversations mainly as problems to be accommodated rather than as foundations worth building from.Within this broader gap, several specific communities are critically underrepresented. Practitioners working in public sector institutions in low and middle income countries — the civil servants, municipal officers, and agency staff who are actually deploying and overseeing AI systems in resource-constrained environments — are almost entirely absent from multilateral discussions despite having the most direct experience of what governance challenges look like at the implementation level. My research across Nairobi, Kigali, Lagos, and Cape Town consistently found that these practitioners have developed sophisticated and contextually grounded responses to governance problems that global frameworks have not yet encountered, let alone resolved.Indigenous communities, whose data, languages, and cultural materials are increasingly implicated in AI systems, are underrepresented not only in governance discussions but in the research that informs those discussions. Workers involved in data production and AI system maintenance — annotators, content moderators, reviewers — are systematically excluded despite being central to how AI systems actually function.Including these voices requires more than translation services and travel grants, though both matter. It requires redesigning the structure of participation so that knowledge produced outside dominant institutional contexts carries genuine weight in shaping outcomes, not merely in illustrating problems that others then propose to solve.

The formats that dominate international governance dialogues — panel presentations, plenary statements, and structured roundtables — are optimised for the exchange of prepared positions, not for genuine deliberation or knowledge production. If the AI Dialogue wants to generate something more than a record of what participants already believed before they arrived in Geneva, it needs to experiment with formats that create different conditions for engagement. Structured practitioner exchanges would be one of the most valuable innovations the Dialogue could introduce. Rather than asking governments to present their AI governance approaches as finished products, the Dialogue could create facilitated sessions where practitioners from different contexts work through a shared governance challenge together — a specific use case, a documented failure, a regulatory ambiguity that multiple jurisdictions are navigating simultaneously. This format surfaces knowledge that formal presentations tend to obscure and creates conditions for genuine learning across contexts. Regional pre-dialogues held before the Geneva meeting would help address the structural inclusion problem. Communities and organisations that cannot send representatives to Geneva can contribute substantively to regional processes whose outputs are formally integrated into the main Dialogue. This is different from consultation: it means the Geneva conversation is explicitly shaped by what happened in Nairobi, Manila, and Bogotá beforehand, not the other way around. Persistent working groups that continue between annual Dialogues would address the continuity problem that affects almost every international governance initiative. Knowledge built in one session currently dissipates before the next. Groups organised around specific implementation challenges — algorithmic accountability in public services, cross-border data governance, capacity building for smaller states — could maintain momentum and produce outputs that accumulate rather than repeat. The Dialogue should also publish all written inputs openly. Transparency about who is contributing and what they are saying is itself a governance practice worth modelling.

The most instructive examples of effective AI governance are not always the most visible ones. Much of the international conversation defaults to the EU AI Act or the OECD AI Principles as reference points, and while both have contributed to the normative landscape, neither represents the full range of what effective governance can look like in practice. Some of the most grounded and contextually coherent approaches are emerging in places that receive comparatively little attention in Geneva-level discussions. Rwanda's approach to digital governance offers a genuinely instructive case. Rather than attempting to regulate AI as a discrete technology, Rwanda has embedded governance considerations into its broader digital infrastructure development, creating accountability mechanisms at the point where AI systems intersect with public service delivery. This integration model - building governance into deployment rather than layering it on afterwards - is more practically effective than the ex-ante regulatory models that dominate Western frameworks, particularly in resource-constrained environments. Kenya's engagement with algorithmic accountability in financial services, particularly in the context of mobile credit scoring systems that affect millions of people with limited formal financial histories, represents another important case. The governance challenges that emerged - opacity of scoring criteria, lack of meaningful recourse, and the reproduction of existing inequalities through proxy variables - were identified and partially addressed through a combination of regulatory guidance and civil society pressure that does not map neatly onto standard regulatory models but produced measurable changes in practice. At the practitioner level, my research has documented governance innovations developed by mid-level public servants and civil society actors that have no formal policy status but are functionally effective. These informal governance practices deserve documentation, analysis, and where appropriate, formalisation. Effective AI governance is already happening in places the Dialogue has not yet looked. The task is to find it, understand it, and build from it rather than assuming it needs to be imported.