Chege Kibathi & Company Advocates, LLP

Success for the first Global Dialogue would not be measured by the elegance of the frameworks debated, but by whether the dialogue honestly confronts a structural problem that most AI governance conversations avoid: the majority of the world's population lives in jurisdictions that lack the technical capacity, institutional independence, and enforcement reach to implement any governance framework, however well designed, in a way that delivers substantive protection. A successful first dialogue would therefore achieve three concrete outcomes. First, it would produce an honest mapping of the governance capacity gap between frontier AI jurisdictions and the rest of the world, not as a development problem to be managed, but as a structural asymmetry that undermines the legitimacy of any global governance framework built without addressing it. Second, it would establish a binding commitment from states with frontier AI capabilities to support the development of independent regulatory infrastructure in Global South jurisdictions, not through technology transfer alone, but through compute access, institutional capacity building, and mutual enforcement frameworks that give developing nations genuine reach over systems operating within their borders. Third, it would move beyond the binary of ex ante regulation versus ex post liability and begin developing a preconditions framework, identifying what structural conditions must exist before governance models can function as intended, and committing resources toward building those conditions in parallel with framework development. A dialogue that produces another non-binding declaration of principles, without addressing who can actually enforce them and against whom, will have failed regardless of how inclusive the process felt.

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

These four priorities reflect a single interconnected argument: that AI governance frameworks, however well designed, cannot deliver substantive protection in jurisdictions that lack the foundational conditions to implement them. AI capacity-building is the threshold priority. Most Global South regulatory authorities face structural constraints: penalty ceilings that cannot reach frontier AI companies, technical staff unable to audit opaque systems, and budgets that cannot sustain meaningful enforcement. Capacity-building is not a supplementary concern. It is the precondition without which every other priority remains aspirational. Interoperability of governance approaches matters because the current landscape produces what might be termed regulatory decoupling, a systematic divergence between formal legal architecture and substantive protective effect. Frameworks designed in Brussels or Washington are being adopted across Africa and Asia in contexts where the underlying assumptions, market leverage, institutional independence, cross-border enforcement reach, do not hold. Interoperability cannot mean harmonisation toward existing frameworks. It must mean developing approaches that function across genuinely different structural conditions.The social, economic, ethical, cultural, linguistic and technical implications of AI demand urgent attention precisely because these implications are not evenly distributed. AI systems deployed across the Global South are predominantly built elsewhere, trained on data that underrepresents local languages and cultural contexts, and governed by legal orders that have no obligation to consider local interests. The communities bearing the greatest risk of AI harm have had the least input into how these systems are designed and governed.Transparency, accountability, and human oversight are foundational to enforcement. A jurisdiction whose regulators cannot compel a foreign AI company to disclose its systems for audit has transparency as a legal right but not a practical reality. The dialogue must address the enforcement dimension of transparency, not merely its declaratory one.

The most significant cross-cutting issue not adequately captured by the listed themes is compute sovereignty, the structural dependency of most Global South nations on foreign-owned infrastructure for the development, deployment, and governance of AI systems. This dependency operates at multiple levels simultaneously. The physical infrastructure, data centres, chips, cloud computing capacity, is concentrated in a small number of jurisdictions. The training data pipelines that determine how AI systems understand and represent the world are controlled by companies headquartered in those same jurisdictions. The legal orders governing data processed by these systems are foreign to the communities most affected by them. And the enforcement mechanisms available to Global South regulators cannot, in practice, reach the entities making the most consequential decisions about how AI systems behave. The consequence is that the ex ante and ex post governance frameworks being debated internationally rest on assumptions, technical capacity to audit, institutional independence from regulated infrastructure, enforcement reach across jurisdictions, that are structurally absent in most developing nations. Governance frameworks imported under these conditions do not simply underperform. They produce a systematic divergence between formal legal protection and substantive effect that no amount of framework refinement can resolve without addressing the underlying structural conditions. The Global Dialogue should therefore treat compute sovereignty as a foundational cross-cutting issue, not a technical matter for specialists, but a governance precondition that determines whether any framework adopted by this dialogue can function as intended across the full diversity of member states it is meant to serve.

Kenya sits at a paradoxical intersection in the global AI governance landscape. It has demonstrated genuine institutional ambition by enacting the Data Protection Act 2019, launching a National AI Strategy 2025–2030, and introducing an Artificial Intelligence Bill 2026 currently before the Senate. Yet this formal architecture coexists with structural conditions that systematically undermine its protective effect. The capacity gap is concrete, not theoretical. The Office of the Data Protection Commissioner, which is Kenya's primary AI governance mechanism by default, given the absence of enacted AI-specific legislation, operates with a statutory maximum penalty capped at KES 5,000,000 or one percent of annual turnover, whichever is lower. This ceiling ensures that the most powerful AI actors operating in Kenya face the least proportionate deterrence. More fundamentally, no Kenyan regulator can compel a foreign AI company to submit its systems for independent audit. The jurisdiction exists on paper. The reach does not. This gap is not abstract. In legal practice, AI tools deployed within Kenyan law firms process sensitive client data through foreign platforms governed by foreign legal orders. Practitioners encounter the governance gap not as a policy question but as an operational reality, improvising workarounds where frameworks should exist, and absorbing risks that no institution is equipped to regulate. The opportunity, however, is equally real. Kenya's demographic profile, its position as East Africa's technology hub, and its relatively developed institutional architecture place it at a unique intersection, capable of demonstrating what meaningful AI governance capacity in the Global South could look like, if the structural preconditions are built rather than assumed. The challenge is that the window for building those preconditions is narrowing faster than current policy cycles can accommodate.

The AI Dialogue's most distinctive contribution to international cooperation would be to reframe the terms of the conversation itself. Current international AI governance efforts, from the EU AI Act to the Bletchley Declaration to the G7 Hiroshima Process, have been designed primarily by and for jurisdictions with frontier AI capabilities. The frameworks they produce reflect the assumptions, risk tolerances, and institutional infrastructures of those jurisdictions. When adopted elsewhere, they frequently generate regulatory decoupling, formal compliance without substantive protective effect, because the underlying preconditions for their function do not exist in most developing nations. The AI Dialogue, convened under UN auspices with universal membership, is uniquely positioned to do what no regional or minilateral forum can: surface the structural asymmetries that existing frameworks systematically obscure, and build international cooperation around addressing those asymmetries rather than papering over them. Concretely, this means the Dialogue should pursue three cooperative functions that existing mechanisms have not adequately addressed. First, it should establish a multilateral compute access framework, ensuring that developing nations can audit, train, and govern AI systems without complete dependence on foreign infrastructure controlled by private actors subject to different legal orders. Second, it should develop mutual enforcement mechanisms that give developing nation regulators genuine cross-border reach, not merely declaratory rights, over AI systems operating within their jurisdictions but governed elsewhere. Third, it should institutionalise the inclusion of practitioner voices from the Global South in governance standard-setting processes, not as consultees after frameworks are designed, but as co-architects from the outset. The Dialogue's added value is not another declaration. It is the only forum with the legitimacy to make these structural corrections binding.

Several existing initiatives provide meaningful foundations for the AI Dialogue to build upon, though each carries limitations that the Dialogue is positioned to address. The African Union's Continental AI Strategy and the Malabo Convention on cybersecurity and data protection represent the most developed regional frameworks for the Global South. However, ratification of the Malabo Convention remains limited, and the AU's AI Strategy lacks enforcement architecture. The Dialogue should engage the AU directly not to duplicate its work but to provide the multilateral legitimacy and resource commitments that regional frameworks cannot generate independently. The OECD AI Principles and the UNESCO Recommendation on the Ethics of AI represent the most widely adopted international normative frameworks. Both are substantively sound but structurally limited, designed through processes dominated by high-income nations and lacking implementation support for jurisdictions without existing governance capacity. The Dialogue should treat these as normative starting points while developing the institutional and financial mechanisms that make them implementable across the full diversity of member states. The Global Digital Compact, adopted in 2024, established commitments on digital cooperation and AI governance that the Dialogue should operationalise rather than renegotiate. Specifically, the GDC's commitments on capacity building and inclusive AI governance provide a mandate the Dialogue can convert into concrete programmes. The added value the Dialogue brings that none of these initiatives can replicate is universality combined with accountability. The UN framework creates the only forum where small, developing, and structurally disadvantaged nations can advance binding commitments on an equal footing with frontier AI states, and where non-compliance carries reputational and institutional consequences that minilateral or regional frameworks cannot impose. The Dialogue should use that leverage deliberately, not decoratively.

The AI Dialogue's legitimacy will ultimately be determined not by who convenes it but by who meaningfully shapes it. Inclusive participation requires structural design choices, not merely aspirational language about openness. For governments, the Dialogue should move beyond position-statement presentations toward structured negotiation sessions with defined deliverables. Governments should arrive having consulted domestic civil society and technical communities, and should be accountable for reporting back on how those consultations informed their positions. For civil society and affected communities, participation cannot be limited to observer status or side events. The Dialogue should establish a formal civil society track with guaranteed speaking time in plenary sessions, dedicated working groups where civil society recommendations are responded to on the record, and transparent documentation of how civil society inputs influenced final outputs. For technical experts, the Dialogue needs a standing advisory mechanism, not a one-time consultation, that provides ongoing technical assessment of governance proposals before they are adopted. Many governance failures stem from frameworks designed without adequate technical input on what is actually enforceable. For practitioners, lawyers, doctors, educators, journalists operating AI tools in their daily work, the Dialogue currently has no meaningful engagement pathway. These are the people who encounter governance gaps as operational realities rather than policy abstractions. Their testimony should be systematically collected and presented as evidence, not anecdote. On format, the Geneva session should be structured around specific problem statements rather than thematic panels. Each problem statement should have a defined output, a recommendation, a commitment, or an identified gap requiring further work, so that the Dialogue produces traceable outcomes rather than general declarations.The measure of inclusive participation is not who was invited. It is whose input changed the outcome.

The most systematically underrepresented voices in global AI governance discussions are not hard to identify, they are the communities bearing the greatest risk of AI harm while having the least influence over how AI systems are designed, deployed, and governed. Practitioners in developing jurisdictions, lawyers, healthcare workers, civil servants, journalists, who interact with AI systems daily but have no pathway into governance processes represent a critical gap. Their knowledge is not theoretical. It is operational evidence of where frameworks fail in practice. The Dialogue should establish a structured practitioner testimony mechanism, collecting and presenting this evidence as primary input rather than illustrative anecdote. Communities whose languages, cultures, and social contexts are underrepresented in AI training data face compounded invisibility, both in the systems themselves and in the governance processes meant to regulate them. Meaningful inclusion requires dedicated representation mechanisms, interpretation support, and governance processes that do not assume English-language participation as the default. Small island states and least developed countries lack the technical capacity and diplomatic bandwidth to engage substantively in multiple simultaneous AI governance processes. The Dialogue should provide dedicated capacity support, not just funding for attendance, but substantive preparation assistance so that these delegations can participate as informed negotiators rather than passive observers. Youth voices are structurally absent from governance processes that will determine the world they inherit. The Dialogue should establish a formal youth track with genuine influence over outputs, not a parallel youth forum whose recommendations are noted and then set aside. Finally, the private sector's participation should be balanced by mandatory disclosure requirements, companies engaging in the Dialogue should be required to disclose relevant commercial interests so that other stakeholders can assess the positions being advanced. Inclusion without influence is not participation. It is the appearance of it.

The traditional UN conference format, prepared statements delivered to a room of delegates checking their phones, is structurally incapable of generating the dynamic engagement that AI governance urgently requires. The Dialogue should deliberately depart from it. Structured adversarial sessions would significantly improve the quality of debate. Rather than parallel presentations of compatible positions, the Dialogue should convene sessions where specific governance proposals are defended by proponents and challenged by designated critics, including critics from civil society and developing nations, with responses required on the record. This format surfaces genuine disagreements rather than allowing them to be papered over by consensus language. Scenario-based deliberation would ground abstract governance debates in concrete operational realities. Rather than debating principles in isolation, delegates would work through specific scenarios, an AI hiring tool deployed across three jurisdictions with different regulatory frameworks, a generative AI system producing medical advice in a language underrepresented in its training data, and identify where existing frameworks fail and what governance responses would be required. This format has proven effective in other international negotiation contexts and would be genuinely novel in AI governance. Asynchronous digital participation, properly designed, could dramatically expand who can meaningfully contribute. Not a comments portal that generates unread submissions, but a structured deliberative platform where inputs are categorised, synthesised, and responded to by working groups, with the synthesis process itself transparent and auditable. Finally, the Dialogue should commission and publicly present independent assessments of AI governance gaps from practitioners in underrepresented regions before each session, so that delegates are confronted with operational evidence of governance failure, not just normative arguments about why better governance is desirable. The format should match the urgency. Business as usual will produce business as usual outcomes.

Several policies and practices offer instructive models for effective AI governance, though each requires contextualisation before adoption elsewhere. Kenya's Data Protection Act 2019 demonstrates that developing nations can establish comprehensive data governance frameworks with limited resources. Its provisions on automated decision-making, data subject rights, and cross-border transfer restrictions are substantively sound. Its limitation, a penalty ceiling that cannot meaningfully deter frontier AI companies, illustrates that framework quality and enforcement capacity are separate problems requiring separate solutions. The lesson is that legislative ambition must be matched by institutional resourcing. The EU AI Act's risk-classification approach, tiering regulatory requirements by the severity of potential harm, offers a structurally sound model for proportionate governance. Its limitation for Global South adoption is the assumption of pre-existing regulatory infrastructure capable of conducting conformity assessments. Adapting this model requires building that infrastructure first, not after. Rwanda's approach to AI and digital infrastructure development, treating compute capacity as sovereign infrastructure requiring deliberate investment rather than market provision, offers an instructive model for smaller developing economies seeking to reduce dependency on foreign-controlled systems. The African Union's data policy framework, while unevenly ratified, demonstrates the value of regional coordination in contexts where individual nations lack the market leverage to negotiate effectively with frontier AI companies. Collective bargaining, whether on data access, audit rights, or penalty frameworks, produces leverage that no individual developing nation can generate independently. Finally, the practice of embedding civil society and practitioner voices in regulatory sandbox design, piloted in several jurisdictions including Kenya's proposed AI sandbox under the 2025 strategy, offers a model for ensuring governance frameworks are tested against operational reality before adoption. The common thread across effective approaches is that they treat governance capacity as infrastructure to be built, not a precondition to be assumed.