São Paulo State University

A first Global Dialogue on AI Governance should be evaluated by whether it produces coordination mechanisms that improve the empirical monitoring and management of real-world AI risks, rather than symbolic commitments. A primary indicator of success would be agreement on a minimum shared taxonomy of AI incidents. Current datasets remain difficult to compare across countries because classifications differ substantially. Even partial convergence around common categories—such as informational, reputational, institutional, economic, and physical harms—would significantly improve cross-national analysis and policy evaluation. Second, the Dialogue should advance regulatory interoperability rather than harmonization. Given heterogeneous institutional capacities and legal traditions, success would consist in aligning principles of transparency, accountability, and traceability for high-risk systems, enabling jurisdictions to coordinate without imposing uniform legislation. Third, a meaningful outcome would be the creation of permanent mechanisms for international evidence sharing on AI incidents. A structured and continuously updated repository—ideally supported by governments, international organizations, and research institutions—would strengthen early-warning capacity and comparative governance research. Fourth, the Dialogue should explicitly prioritize observed risks over speculative scenarios. Empirical evidence suggests that many current incidents are reputational and informational, especially involving misinformation and institutional trust. Governance agendas that neglect these risks remain misaligned with the present risk landscape. Finally, success requires the establishment of a continuing institutional process, including technical working groups, monitoring indicators, and an implementation timeline. Without these elements, the Dialogue would remain declaratory rather than operational.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights
• Interoperability of governance approaches

My selection reflects priorities that are both empirically grounded and institutionally actionable in the current stage of global AI governance. Safe, secure and trustworthy AI is a central priority because the expansion of AI systems across public and private sectors increases exposure to operational failures, misuse, and unintended harms. Strengthening safety and security standards is essential for reducing real-world incidents and improving the reliability of deployed systems, particularly in high-impact domains. Interoperability of governance approaches is urgent given the rapid proliferation of national regulatory frameworks. Without coordination mechanisms, fragmentation may reduce policy effectiveness and complicate cross-border monitoring of AI risks. Promoting interoperability enables jurisdictions to cooperate while preserving regulatory flexibility and institutional diversity. Protection and promotion of human rights remains a foundational requirement for legitimate AI governance. Existing evidence shows that AI-related incidents frequently involve discrimination, privacy violations, and risks to freedom of expression. Ensuring that governance frameworks are anchored in internationally recognized rights is therefore necessary to align technological development with democratic norms and social trust. Transparency, accountability, and human oversight are essential for making governance frameworks operational rather than declaratory. Clear responsibility structures, documentation standards, and oversight mechanisms enable auditing, contestability, and corrective intervention when harms occur. These elements are particularly important in high-stakes contexts such as public administration, security, and digital information ecosystems. Together, these priorities reflect a strategy focused on strengthening institutional capacity to monitor risks, improve cross-jurisdictional coordination, and ensure that AI deployment remains consistent with established legal and normative principles.

The listed themes address core governance concerns, several cross-cutting and emerging issues remain insufficiently captured. First, there is a need for stronger emphasis on empirical monitoring of AI incidents. Current governance discussions often rely on forward-looking risk scenarios rather than systematically documented evidence of observed harms. Building internationally comparable incident-reporting infrastructures would improve early-warning capacity, regulatory calibration, and policy evaluation. Second, data governance asymmetries across countries represent a growing structural issue. Differences in access to datasets, computational resources, and technical expertise risk reinforcing global inequalities in AI development and oversight. Addressing these asymmetries is essential for ensuring meaningful participation by lower- and middle-income countries in global governance processes. Third, measurement standards for risk classification and impact assessment remain underdeveloped. Without shared methodological baselines, comparisons across jurisdictions and sectors remain limited. Establishing interoperable indicators for evaluating harms, exposure, and mitigation effectiveness would significantly strengthen international coordination. Fourth, the institutional capacity of public administrations to implement AI oversight deserves greater attention. Many regulatory frameworks assume monitoring capabilities that are not yet widely available. Investment in technical expertise, audit infrastructures, and supervisory tools is therefore a prerequisite for effective governance. Finally, the governance implications of informational and reputational harms, particularly those associated with synthetic media and large-scale misinformation, remain underprioritized relative to other categories of risk. These harms already affect political institutions and public trust and should be treated as central, rather than secondary, elements of the global AI risk landscape.

In Brazil and, more broadly, in Latin America, governance gaps in safe, secure and trustworthy AI, interoperability of regulatory approaches, human rights protection, and transparency and accountability mechanisms are shaping both the risks and opportunities associated with AI adoption. A central challenge concerns institutional capacity for oversight. While Brazil has advanced regulatory discussions and sectoral initiatives, implementation capabilities remain uneven across public agencies. This affects the ability to audit high-risk systems, monitor incidents, and enforce accountability requirements in practice. A second challenge is regulatory fragmentation across jurisdictions. Differences between emerging national frameworks in Latin America and more consolidated approaches in regions such as the European Union create uncertainty for cross-border cooperation, compliance alignment, and data governance. Limited interoperability reduces the effectiveness of regional coordination on shared risks such as misinformation and automated decision-making in public services. There are also important human rights risks, particularly in relation to biometric surveillance, automated eligibility systems in social policies, and large-scale informational harms. These areas require clearer safeguards to prevent discriminatory outcomes and to ensure procedural fairness in high-impact decisions affecting vulnerable populations. At the same time, significant opportunities are emerging. Brazil has strong academic capacity, an active civil society, and increasing participation in international AI governance discussions. These factors create favorable conditions for contributing to evidence-based governance approaches, especially through incident monitoring, comparative policy evaluation, and the development of context-sensitive regulatory tools adapted to middle-income countries. Strengthening transparency standards, improving technical expertise within regulatory institutions, and expanding international cooperation mechanisms could position the region not only as a policy recipient but as an active contributor to global AI governance frameworks.

The AI Dialogue can play a meaningful role in advancing international cooperation by functioning as a coordination platform between existing regulatory initiatives rather than as a venue for negotiating a single global framework. First, it can support the development of shared technical baselines, including common taxonomies of AI risks, incident-reporting standards, and evaluation metrics. These elements are prerequisites for comparability across jurisdictions and for evidence-based policymaking at the global level. Second, the Dialogue can promote interoperability among governance regimes by aligning principles and risk-classification approaches across national and regional frameworks. Given the rapid expansion of heterogeneous regulatory models, coordination mechanisms are essential to reduce fragmentation and facilitate cross-border oversight of high-risk systems. Third, it can strengthen international information-sharing infrastructures. Regular exchange of empirical evidence on AI incidents, mitigation strategies, and regulatory outcomes would improve early-warning capacity and enable governments to respond more effectively to emerging risks. Fourth, the Dialogue can help expand inclusive participation from developing and middle-income countries, whose perspectives are often underrepresented in global technology governance. Supporting technical cooperation, capacity-building initiatives, and access to shared monitoring resources would make governance processes more representative and operationally effective. Finally, the Dialogue can contribute by establishing continuity mechanisms, such as technical working groups, monitoring indicators, and implementation roadmaps. These structures transform high-level commitments into sustained cooperation over time. In this sense, the Dialogue's main value lies not in producing binding agreements, but in creating institutional conditions for coordination, comparability, and cumulative learning across national AI governance efforts.

The AI Dialogue should build on existing multilateral and technical initiatives that already structure international cooperation in AI governance, while addressing coordination gaps between them. A first reference point is the OECD AI Principles and the OECD AI Policy Observatory, which provide widely adopted normative baselines and comparative policy monitoring tools. The Dialogue could extend their impact by promoting broader participation from countries not currently engaged in OECD-led processes. Second, the UNESCO Recommendation on the Ethics of Artificial Intelligence offers a globally negotiated framework grounded in human rights and institutional safeguards. The Dialogue could support implementation by facilitating technical exchange on operationalizing these principles across regulatory systems. Third, the Global Partnership on Artificial Intelligence (GPAI) has produced important work on responsible AI, data governance, and future-of-work impacts. However, its membership structure is limited. The Dialogue could serve as a more inclusive platform linking GPAI outputs with broader UN processes. Fourth, regional regulatory developments such as the European Union AI Act are shaping emerging global standards, particularly in risk classification, compliance obligations, and the use of regulatory sandboxes to support supervised experimentation with innovative systems. The Dialogue could help disseminate best practices from these sandbox environments, enabling countries with different institutional capacities to test governance tools in controlled settings. The added value of the AI Dialogue lies in institutional coordination rather than duplication. It can connect normative frameworks, technical standards, regulatory experimentation mechanisms, and evidence-sharing initiatives into a more coherent global architecture. In particular, it can support common incident-reporting practices, strengthen participation from underrepresented regions, and promote comparability across governance approaches without requiring legal harmonization.

Different stakeholders can contribute to the AI Dialogue by performing complementary roles in evidence production, standard-setting, implementation, and oversight. The effectiveness of the Dialogue will depend on whether these roles are clearly structured rather than treated as purely consultative participation. Governments should provide regulatory experience, national risk assessments, and information on implementation constraints. Their contribution is essential for aligning governance principles with institutional feasibility and legal accountability structures. International organizations can support coordination across jurisdictions by developing shared taxonomies, monitoring indicators, and reporting infrastructures. They are particularly well positioned to ensure comparability across national approaches and continuity between Dialogue sessions. Academic institutions should contribute empirical research on AI incidents, evaluation methodologies, and impact measurement frameworks. Evidence-based governance requires systematic integration of independent research rather than ad hoc expert consultation. Private-sector actors can provide technical knowledge on system design, deployment risks, and compliance challenges. Their participation is especially important for testing audit mechanisms, documentation standards, and regulatory sandbox environments. Civil society organizations play a critical role in identifying rights-related risks, monitoring social impacts, and ensuring representation of affected communities, particularly in high-stakes applications such as public services and digital information ecosystems. In terms of structure, the Dialogue should include permanent technical working groups, standardized reporting mechanisms, and a multi-year implementation roadmap. Regular publication of monitoring indicators and incident summaries would strengthen transparency and policy learning. A rotating regional consultation track would also improve participation from underrepresented countries and help align global principles with diverse institutional contexts.

Global discussions on AI governance continue to underrepresent perspectives from low- and middle-income countries, subnational public-sector implementers, labor-market stakeholders, data-constrained research communities, and populations directly affected by informational harms such as misinformation and synthetic media. As a result, many governance frameworks reflect the priorities and institutional capacities of a limited group of technologically advanced jurisdictions, reducing their applicability elsewhere. These gaps can be addressed by expanding regional consultation tracks, supporting national and cross-country incident-reporting infrastructures, enabling access to shared datasets and research tools, and ensuring sustained participation of these actors in technical working groups rather than limiting them to symbolic consultation processes. Such inclusion would improve the empirical grounding, legitimacy, and operational effectiveness of global AI governance efforts.

Meaningful engagement in the AI Dialogue would benefit from formats that prioritize technical exchange, comparability of evidence, and continuity between meetings rather than one-time plenary discussions. One effective format would be thematic technical working groups organized around specific governance challenges such as incident reporting, risk classification, and audit methodologies. These groups could produce shared taxonomies, indicators, and implementation guidance between Dialogue sessions. A second format would involve structured policy labs or regulatory sandbox exchanges, where participating countries present ongoing experiments with oversight tools, documentation standards, and supervised deployment environments. This would allow jurisdictions with different institutional capacities to learn from practical regulatory experience rather than abstract principles alone. Third, comparative evidence sessions based on standardized country case submissions could improve cross-national learning. Short technical reports on incident trends, regulatory pilots, or evaluation results would strengthen the empirical foundation of discussions and support cumulative policy learning. Fourth, multi-stakeholder scenario exercises could help participants assess governance responses to realistic risk situations, including cross-border misinformation campaigns or failures in automated public-sector systems. These exercises would clarify institutional responsibilities and coordination needs. Finally, establishing a permanent monitoring and reporting track linked to the Dialogue would ensure continuity. Regular publication of indicators, incident summaries, and implementation progress would transform the Dialogue into an iterative governance process rather than a periodic consultation forum.

Regulatory sandboxes represent one of the most practical instruments currently available to support effective and adaptive AI governance. They allow governments to test oversight mechanisms in controlled environments while enabling innovation under supervisory conditions. By facilitating structured experimentation with documentation requirements, risk classification procedures, audit methods, and monitoring tools, sandboxes help regulators evaluate policy feasibility before large-scale implementation. An important example is the sandbox framework incorporated into the European Union AI Act, which encourages national authorities to create supervised testing environments for high-risk AI systems. These mechanisms reduce uncertainty for developers while strengthening regulatory learning and institutional capacity. Similar initiatives in multiple jurisdictions demonstrate that sandbox environments can support iterative policy design rather than static compliance models. Regulatory sandboxes are particularly valuable for countries with evolving governance infrastructures, as they allow gradual alignment with international standards without requiring immediate legal harmonization. They also create opportunities for collaboration between regulators, academic researchers, and private-sector developers, improving transparency and accountability during early deployment stages. In addition, cross-border cooperation between sandbox initiatives could strengthen interoperability of governance approaches by enabling the comparison of testing methodologies, evaluation indicators, and supervisory practices. Establishing shared reporting formats across sandbox programs would further contribute to the development of internationally comparable evidence on AI risks and mitigation strategies. For these reasons, regulatory sandboxes provide a concrete and scalable mechanism for translating high-level governance principles into operational regulatory practice while supporting innovation under structured oversight conditions.