Hyper Transformation LLC

A successful first Global Dialogue on AI Governance would produce actionable guidance that helps translate widely shared principles into implementable governance mechanisms across diverse institutional and regional contexts. Current global discussions have made significant progress in defining safe, trustworthy and rights based AI. However, they remain largely focused on evaluating system outputs, often overlooking a more fundamental upstream condition: whether AI systems are capable of meaningfully engaging with the populations they are intended to serve. In many regions, particularly across the Global South, informality, limited documentation and uneven access to digital infrastructure mean that large segments of the population do not generate the types of data that AI systems are designed to interpret. As a result, governance frameworks may appear robust while failing to detect a structural limitation, namely that some individuals are not being meaningfully evaluated at all. Addressing this gap requires complementing existing approaches with governance mechanisms that assess evaluability as part of system design. This includes integrating coverage considerations into risk and impact assessments, incorporating data inclusion criteria into public procurement, and strengthening institutional capacity to align AI systems with real socio economic conditions. A strong outcome of the Dialogue would therefore be a shared understanding that effective AI governance must operate not only at the level of outputs, but also at the level of design assumptions. This would enable more inclusive, context aware and implementable governance approaches, while strengthening trust and coherence across global efforts.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight

The selected priorities reflect the need to strengthen the institutional conditions under which AI governance can be effectively implemented, particularly in contexts where system design assumptions and real world conditions diverge. Safe, secure and trustworthy AI provides the necessary baseline for system reliability, but on its own it does not ensure that governance mechanisms can operate effectively across different environments. This is why capacity building is essential, not only in technical terms, but in enabling public institutions to interpret, adapt and enforce governance frameworks in conditions of uneven data availability and institutional maturity. The inclusion of social, economic and technical implications is critical to ensure that governance frameworks account for how AI systems interact with existing structures of inequality, informality and limited digital infrastructure. Without this perspective, governance risks being calibrated to idealized conditions that do not reflect the environments in which systems are deployed. Transparency, accountability and human oversight remain central, but their effectiveness depends on their ability to engage with how systems are designed and implemented in practice. In contexts where institutional capacity and data environments vary significantly, these mechanisms must be supported by governance approaches that can operate under constraint, rather than assuming standardized conditions. Taken together, these priorities enable a shift from abstract alignment around principles toward governance approaches that are institutionally grounded, context aware and capable of addressing implementation challenges across diverse settings.

A cross cutting issue that remains insufficiently captured by the listed themes is the problem of "non evaluation" as a structural limitation of current AI systems. Most existing governance frameworks are designed to address risks that arise once a system produces an output. They focus on ensuring that outputs are safe, fair, transparent and accountable. This architecture implicitly assumes that individuals are already within the system's evaluative scope, and that governance can be exercised by interrogating outputs. In practice, however, this assumption does not hold across many real world contexts. In large parts of the Global South, as well as within underserved populations in high income countries, economic and social activity is not consistently translated into the types of structured data that AI systems are designed to interpret. Informality, limited documentation, fragmented data infrastructures and unequal digital access create conditions in which individuals are not only at risk of biased evaluation, but of not being meaningfully evaluated at all. This introduces a distinct governance problem. Exclusion does not occur as a deviation within the system, but as a consequence of how the system defines what counts as evidence. In such cases, governance mechanisms that rely on auditing outputs are structurally unable to detect the issue, because the absence of evaluation produces neither a contestable output nor a visible error. As a result, non evaluation operates as an upstream condition that cuts across all thematic areas, including human rights, capacity building and trustworthy AI, while remaining largely invisible within current frameworks. It reflects a deeper misalignment between system design assumptions and the socio economic realities of deployment contexts. Addressing this gap requires expanding AI governance beyond output based oversight toward the examination of design stage decisions, particularly the definition of admissible evidence and the coverage of evaluation across populations. This would enable governance frameworks to identify where systems fail not because they perform poorly, but because they are not equipped to engage with the populations they are intended to serve. Recognizing non evaluation as a governance issue would strengthen the Dialogue by ensuring that inclusion is treated as a condition of system design and not only as an outcome to be measured after deployment.

Across Latin America, AI governance gaps are not only regulatory, but structural, and they are already affecting how systems are deployed, interpreted and trusted across key sectors. While there has been significant progress in adopting national AI strategies and ethical frameworks, implementation capacity remains uneven, particularly within public institutions. This creates a persistent disconnect between high level commitments to safe, trustworthy and rights based AI and the realities of procurement, deployment and oversight. A central challenge is that many governance approaches assume the availability of high quality, structured and interoperable data. In practice, however, large segments of the population operate within informal or semi formal systems, where economic and social activity is not consistently captured by digital infrastructures. This leads to a systemic misalignment between AI system design and the socio economic conditions in which systems operate. The impact is particularly visible in sectors such as finance, social protection, health and public services, where governance mechanisms focused on fairness, transparency and accountability at the level of outputs often fail to detect a more fundamental limitation: whether systems are capable of meaningfully evaluating the populations they are intended to serve. As a result, exclusion can occur without being visible as bias or error, limiting both effectiveness and trust. At the same time, these challenges create important opportunities. Governments can strengthen AI governance by integrating context aware approaches into procurement, risk assessment and capacity building, and by requiring greater transparency around system assumptions and data dependencies. Regional collaboration and engagement with international frameworks also provide a foundation for developing more adaptable governance models. The key opportunity is to translate this momentum into governance mechanisms that are not only technically robust, but also operationally viable and aligned with the diversity of real world conditions across the region.

The AI Dialogue can play a critical role in advancing international cooperation by addressing one of the central weaknesses of the current AI governance landscape: the gap between normative convergence and operational divergence. Many international frameworks now share similar language around safety, trustworthiness, human rights, transparency and accountability. However, countries and institutions differ significantly in their capacity to interpret, implement and enforce these principles. The result is a fragmented form of alignment, where agreement exists at the level of values, but diverges in practice. The Dialogue can strengthen cooperation by acting as a coordination and translation mechanism. Its role should not be to produce new principles, but to clarify how existing frameworks can be applied, adapted and combined across different institutional contexts. This includes identifying areas of convergence and conflict, and generating shared understanding of how governance approaches perform under real deployment conditions. A key contribution would be to develop comparative implementation pathways, documenting how different countries operationalize similar principles under varying constraints. This would enable more practical forms of cooperation, allowing countries to learn from each other's approaches rather than attempting to replicate models that may not fit their context. The Dialogue can also support cooperation by legitimizing diversity in governance approaches, while maintaining coherence at the level of shared objectives. This would strengthen interoperability without requiring uniformity. In this sense, the AI Dialogue can serve as a bridge between global frameworks and local governance realities, enabling cooperation that is not only aligned in principle, but effective in practice.

The AI Dialogue should build on existing initiatives that have established the normative foundations of AI governance, including the OECD AI Principles, UNESCO's Recommendation on the Ethics of AI, relevant United Nations processes, the Global Digital Compact, and regional regulatory developments such as the European Union AI Act. These frameworks have created important convergence around safety, trustworthiness, human rights, transparency, accountability and inclusion. The remaining challenge, however, lies in how these principles are interpreted and implemented across diverse institutional contexts. Many countries face overlapping frameworks, uneven implementation capacity and uncertainty about how to operationalize global standards in practice. The added value of the AI Dialogue should therefore be to function as a coordination and translation mechanism. Rather than creating new principles, it should map how existing frameworks align, identify where they diverge in implementation, and generate practical guidance on how they can be applied under different institutional conditions. A key contribution would be to develop comparative implementation insights, showing how similar governance principles are operationalized across countries with varying levels of data infrastructure, regulatory maturity and institutional capacity. This would allow policymakers to adapt governance approaches rather than replicate them. The Dialogue can also add value by surfacing shared operational gaps that are not fully addressed by existing initiatives, particularly in contexts where systems interact with heterogeneous data environments and complex socio economic realities. In this way, the AI Dialogue can strengthen international cooperation by connecting existing frameworks into a more coherent and usable governance ecosystem, enabling countries to move from normative alignment toward practical and context sensitive implementation.

Different stakeholders should contribute according to the type of knowledge they are best positioned to provide. Governments should share implementation experience, including regulatory constraints, procurement challenges and public sector deployment risks. The private sector should provide insight into system design choices, operational limitations, data dependencies and deployment trade offs. Civil society should document rights impacts, exclusion risks and lived experiences of affected communities. Academia and technical experts should contribute analytical methods, evaluation tools and evidence on system performance. International organizations should help synthesize these inputs and connect them to existing governance frameworks. The Dialogue should be structured around concrete governance problems rather than broad stakeholder statements. Each session should bring different actors together around specific questions, such as public sector procurement, accountability mechanisms, human oversight, data governance, capacity building or deployment in low resource contexts. The format should include three layers: first, short evidence submissions before each session; second, structured working sessions focused on producing practical outputs; and third, public synthesis documents showing how stakeholder input was used. The Dialogue should also include regional tracks, so that perspectives from Latin America, Africa, Asia and small island states are not treated as secondary to global discussions. Finally, it should use iterative feedback cycles, allowing stakeholders to review and refine draft outputs before final recommendations are adopted. This would make the Dialogue more than a consultation. It would turn it into a structured process for translating diverse expertise into actionable AI governance guidance.

Global discussions on AI governance currently underrepresent three groups in particular. First, communities living and working outside formal data infrastructures: informal workers, undocumented or weakly documented populations, rural communities, microentrepreneurs, migrants, and people whose economic activity is not consistently captured through digital or institutional records. Second, public institutions and practitioners from the Global South, especially those responsible for procurement, social protection, financial inclusion, health, education and digital transformation. These actors often face the practical consequences of AI deployment, but are less represented in agenda setting. Third, local civil society organizations, women's groups, community based organizations and researchers working directly with affected populations. Their knowledge is often treated as contextual or anecdotal, rather than as evidence that should shape governance design. They should be included through more than open consultations. The Dialogue should create structured channels for these groups to influence problem definition, not only provide feedback on predefined agendas. This could include regional evidence submissions, partnerships with local institutions, targeted consultations with affected communities, and expert panels that include implementation practitioners from public agencies, not only global policy specialists. Inclusion should also extend to the evidence used in governance. Context specific knowledge, qualitative evidence and locally grounded implementation experience should be incorporated into policy outputs alongside technical and legal expertise. Without this, global AI governance may become more geographically diverse in appearance while still being shaped by the assumptions of highly formalized, data rich environments.

Innovative engagement formats for the AI Dialogue should move beyond discussion based participation toward mechanisms that generate structured, comparable and policy relevant evidence across different contexts. One approach would be to introduce "governance stress testing labs," where participants evaluate how existing AI governance frameworks perform under specific real world conditions, including low data availability, high informality, or limited institutional capacity. Rather than hypothetical discussion, these labs would require participants to apply governance models to constrained scenarios, revealing where assumptions break down. A second format could involve "evaluability mapping exercises," where stakeholders identify which populations, sectors or use cases fall outside the effective scope of current AI systems. This would make visible forms of exclusion that are not captured by standard metrics such as bias or accuracy, and would generate inputs that can directly inform policy design. The Dialogue could also incorporate "policy prototyping cycles," where small, diverse groups develop and iteratively refine governance interventions based on real deployment constraints. These prototypes would be tested against different regional conditions and revised through structured feedback loops. Finally, creating a shared repository of implementation cases, including both successful and failed deployments, would enable cumulative learning across regions. This would shift the Dialogue from a one time consultation into an evolving evidence base that supports more grounded and adaptive AI governance.

Several existing approaches offer valuable building blocks for effective AI governance, particularly when they move beyond principles and address implementation conditions. Risk management frameworks such as the NIST AI Risk Management Framework are important because they translate high level commitments into operational processes. Their strength lies in structuring governance as an ongoing practice, rather than a one time compliance exercise. Normative frameworks, including the OECD AI Principles and UNESCO's Recommendation on the Ethics of AI, have established a shared baseline around human rights, trustworthiness and accountability. Their impact, however, depends on their ability to be integrated into institutional processes such as procurement, oversight and capacity building. Regulatory approaches such as the European Union AI Act further demonstrate how governance can be operationalized through risk classification, documentation requirements and lifecycle obligations. These mechanisms provide a useful reference for structuring accountability, even in contexts where regulatory models differ. At the implementation level, public procurement stands out as one of the most effective governance levers. Requiring transparency on system purpose, data dependencies, intended users, limitations and human oversight before deployment can significantly improve accountability. A critical area for advancement is the evolution of impact assessment practices. Current approaches often focus on bias, privacy and safety, but less on whether systems are capable of functioning meaningfully within the contexts in which they are deployed. Expanding assessment frameworks to include these conditions would strengthen their relevance. Taken together, these examples suggest that effective AI governance depends not on any single framework, but on the alignment between principles, institutional processes and real world deployment conditions.