Brazillian Ministry of Health

The first Global Dialogue on AI Governance will succeed if it produces not a declaration of principles — the field is saturated with those — but a shared diagnostic framework capable of distinguishing why AI governance fails, not merely that it does. Current multilateral discussions conflate structurally distinct failure modes: execution failures, where governance signals exist but are not escalated; and constitutional failures, where equitable outcomes are architecturally impossible because the normative constraints were never embedded in the system's computational substrate. These require fundamentally different remediation pathways, and no existing framework — NIST AI RMF, ISO 42001, or the EU AI Act as currently operationalized — formally distinguishes between them. A successful Dialogue would therefore produce: (1) a shared typology of AI governance failures that moves beyond checklist-based compliance toward architectural accountability; (2) a commitment to operationalizing human oversight not as a procedural checkbox (as in EU AI Act Article 14) but as a computationally verifiable guarantee — what I term compliance-by-construction as distinct from compliance-by-verification; (3) an agreed methodology for comparing governance frameworks across jurisdictions that respects institutional heterogeneity while identifying structural convergences — enabling genuine interoperability rather than superficial harmonization; and (4) concrete capacity-building pathways that transfer not only technical tools but the normative reasoning architectures that make those tools accountable to democratic mandates. For the Global South, success means a Dialogue that does not reproduce the asymmetry of the current landscape, in which governance frameworks are produced in Washington and Brussels and imposed through market access conditions. AI governance must be built as a genuinely multilateral normative infrastructure — one in which Brazil, the African Union, and ASEAN nations co-produce the frameworks that govern systems affecting their populations.

• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches

My four priorities reflect a systemic diagnosis developed through comparative empirical research across 34 governance documents in 10 jurisdictions, combined with applied cybernetic architecture work in Brazil's public health and agrarian development sectors. Transparency, accountability, and human oversight is the most urgent priority because current frameworks mandate oversight without providing computational mechanisms to operationalize it. Article 14 of the EU AI Act requires "human oversight" for high-risk systems but is silent on implementation architecture. This silence allows compliance to reduce to documentation rather than engineered guarantees. Genuine accountability requires distinguishing execution failures from constitutional failures - a distinction no current multilateral framework formalizes. Interoperability of governance approaches is strategically indispensable. The proliferation of national AI regulations - Brazil's PL 2338/2023, the EU AI Act, executive orders in the United States, emerging ASEAN frameworks - creates jurisdictional fragmentation that disproportionately burdens Global South actors who lack the institutional capacity to simultaneously comply with multiple conflicting normative regimes. Interoperability must be architectural, not merely rhetorical: it requires shared ontologies, compatible audit schemas, and mutual recognition of compliance mechanisms. Safe, secure and trustworthy AI must be grounded in formal verification, not aspirational language. Trustworthiness is only meaningful when it is computationally demonstrable - when normative constraints are embedded as structural invariants of the system, not as external filters applied post-inference. Social, economic, ethical, cultural, linguistic and technical implications anchors the technical discussion in its material consequences. In public infrastructure deployments - healthcare, social protection, food security - the populations most exposed to AI system failures are precisely those with the least representation in the governance processes that produced the frameworks they operate under. Any dialogue that ignores this structural asymmetry will reproduce it.

Two cross-cutting issues are structurally absent from the listed themes and merit explicit recognition. First: the distinction between compliance-by-verification and compliance-by-construction. All thematic areas listed presuppose a verification model of governance - frameworks describe what should be controlled, and external actors verify whether it is. This model does not scale. As the number of high-risk AI systems grows (the European Commission projected over 50,000 requiring conformity assessment by 2027), the verification burden on national competent authorities becomes computationally and institutionally intractable. An emerging alternative - which I term compliance-by-construction - embeds normative constraints as architectural invariants of the AI system itself, shifting the regulator's role from continuous verification to one-time architecture certification. This architectural turn is not captured by any of the listed themes, yet it represents the most consequential frontier in AI governance engineering. Second: ontological sovereignty and the governance of normative representation. When AI systems are deployed in public infrastructure, they inevitably perform classificatory operations: who counts as "food insecure," "medically urgent," or "economically vulnerable." These classifications are not neutral technical choices - they are normative acts that can erase or preserve the legal categories through which marginalized populations access rights. I define this as the problem of ontological sovereignty: the guarantee that legally protected categories (indigenous communities, persons with disabilities, pregnant women in risk conditions) cannot be collapsed into generic statistical categories by gradient optimization. No existing governance framework provides formal mechanisms to enforce ontological sovereignty. The Global Dialogue should address this directly, particularly for AI systems deployed in social protection, healthcare, and food security contexts in the Global South, where the populations at risk are precisely those whose categorical identities were hardest won through decades of rights-based advocacy.

Brazil operates one of the world's largest universal public health systems — the SUS, serving over 190 million people under constitutional guarantees of universality, integrality, and equity. AI systems are rapidly being deployed across this infrastructure, yet the governance frameworks available to regulate them are structurally inadequate to the failures they produce. Two documented cases illustrate the pattern. In January 2021, epidemiological data distributed across DATASUS (SINAN, SIH, CNES) contained all signals necessary to anticipate the catastrophic oxygen supply collapse in Manaus weeks before it occurred. The governance failure was one of execution: the convergent signal existed but was never formally escalated through institutional channels. This is a failure that transparency and accountability frameworks, as currently designed, cannot prevent — because no existing mechanism formalizes the obligation to act when data-derived signals reach critical thresholds. The second pattern is constitutional: AI systems trained on historical SUS data learn to replicate and amplify the structural geographic concentration of specialized healthcare, directing algorithmic recommendations toward already well-served Southeast centers and away from the North and Northeast. This is not a bias that can be corrected by monitoring outputs — it is embedded in the system's foundational architecture. No current multilateral governance framework distinguishes this from the Manaus failure, and therefore no framework provides adequate remediation guidance for either. The broader opportunity is equally significant. Brazil's public data infrastructure — DATASUS, CadÚnico, CAF/MapaSAN — constitutes one of the richest longitudinal datasets on population welfare in the Global South. If governed with architectural accountability rather than procedural compliance, this infrastructure could become a model for how emerging economies build AI systems that are genuinely answerable to constitutional mandates. The governance gap is real, but so is the institutional capacity to close it — if multilateral frameworks create space for non-Northern governance architectures to be taken seriously.

The AI Dialogue arrives at a moment of structural paradox in global AI governance: normative consensus has never been broader, yet the gap between declared principles and operational accountability has never been wider. Comparative analysis of 34 governance documents across 10 jurisdictions — applying fuzzy-set Qualitative Comparative Analysis to seven structural dimensions — confirms that all existing frameworks converge on procedural requirements (documentation, impact assessments, oversight mandates) while sharing a structurally absent third layer: the capacity to verify whether normative objectives are computationally achievable in deployed systems. The Dialogue's distinctive role is not to produce one more declaration, but to make this diagnostic shift — from aspirational governance to architectural accountability — the organizing principle of international cooperation. Concretely, the Dialogue can advance cooperation in three ways that no existing multilateral mechanism has yet achieved. First, it can establish a shared taxonomy of AI governance failures that distinguishes execution failures from constitutional failures — a distinction with direct implications for how international frameworks assign responsibility, mandate remediation, and structure mutual recognition agreements. Second, it can broker agreement on minimum architectural standards for human oversight that are computationally specified, not merely rhetorically required — moving beyond EU AI Act Article 14's silence on implementation to internationally recognized technical criteria. Third, and most consequentially, it can create deliberative space where governance architectures developed outside Washington and Brussels receive institutional legitimacy. The current landscape reproduces an epistemic asymmetry in which the Global South adopts governance frameworks whose categories were defined without their populations in view. A Dialogue that genuinely operates under the UN's universality mandate can interrupt this pattern — not by imposing uniformity, but by establishing interoperability protocols that allow nationally rooted governance frameworks to achieve mutual recognition without requiring normative capitulation.

Several existing mechanisms provide essential infrastructure the Dialogue should build upon rather than duplicate. The OECD AI Principles (updated 2024) and the associated OECD AIHEG (AI and Health Expert Group) represent the most technically grounded multilateral normative consensus currently available, and their sectoral working groups — particularly in health — have generated comparative evidence on governance gaps that should directly inform Dialogue deliberations. The UNESCO Recommendation on the Ethics of AI (2021) established the broadest intergovernmental normative foundation. The Global Partnership on AI (GPAI) has produced substantive working group outputs on responsible AI and data governance. The ITU AI for Good platform provides technical capacity-building infrastructure, particularly relevant for developing nations. The added value the Dialogue can uniquely provide is the connection between these normative layers and a third layer that none of them has yet operationalized: the translation of agreed principles into computationally verifiable governance requirements. The OECD Principles describe what AI governance should achieve; the EU AI Act specifies procedural compliance; neither provides a mechanism for certifying that the normative objectives are architecturally embedded in deployed systems. The Dialogue, under UN auspices with the legitimacy this confers, is positioned to convene the technical, legal, and institutional communities necessary to develop this third layer as a shared international infrastructure — not owned by any single regulatory jurisdiction. Two specific connections are strategically important. The BRICS+ digital governance workstream, currently developing shared frameworks for AI in public services, represents a natural partner for piloting architectural interoperability standards in health, food security, and social protection contexts. The G20 AI Principles track, in which Brazil held the presidency in 2024, has already produced commitments on trustworthy AI that the Dialogue can operationalize. The moment for connecting these parallel tracks into a coherent international architecture is precisely now.

The AI Dialogue's structural challenge is not a deficit of interested stakeholders but a deficit of deliberative architecture — the absence of a format capable of converting heterogeneous contributions into actionable governance outputs rather than ceremonial declarations. Meaningful multi-stakeholder participation requires three design commitments. First, the Dialogue must distinguish between consultative and constitutive participation. Inviting civil society to observe plenary sessions is consultative; creating working groups where civil society organizations, technical researchers, and affected community representatives co-draft the diagnostic frameworks against which AI governance proposals are evaluated is constitutive. The latter produces governance outputs that carry the epistemic weight of plural authorship. Technical researchers should contribute not only as expert witnesses but as co-designers of the evaluation criteria — including the criteria for what counts as adequate human oversight, which is currently undefined in every major multilateral framework. Second, the format must be structured to prevent the well-documented tendency of multistakeholder processes toward capture by the best-resourced participants. Large technology companies have dedicated policy teams whose full-time occupation is engaging multilateral governance processes; most civil society organizations, academic researchers, and government delegations from developing nations do not. Structured time-allocation guarantees, pre-session working document access in multiple languages, and explicit rapporteur roles for underrepresented constituencies are minimum procedural requirements. Third, governments should participate not only as regulators but as operators. Brazil, India, South Africa, and other states that run large-scale AI deployments in public health, social protection, and education have operational knowledge of governance failure that no private sector actor possesses. A track explicitly designed for government-as-operator experiences — distinct from government-as-regulator positions — would surface the implementation realities that current frameworks systematically miss: what actually happens when an AI system interacts with a constitutional mandate in a resource-constrained public institution.

The most consequential underrepresentation in global AI governance discourse is structural, not incidental, and it operates at three levels. The first is geographic-institutional. Governance frameworks are produced in Washington, Brussels, London, and Geneva by organizations whose staff, funding, and epistemic reference points are overwhelmingly Northern. The Global South is invited to adopt, adapt, and implement these frameworks — not to co-produce them. This is not merely a matter of fairness; it is a matter of validity. AI systems deployed in Brazil, Nigeria, Bangladesh, and Indonesia operate within normative contexts — constitutional mandates, customary law, communal land tenure systems, indigenous health categories — that Northern-designed frameworks do not represent and cannot govern adequately. Boaventura de Sousa Santos identified this as epistemicide: the suppression of knowledge systems by the imposition of external ontological categories. Governance processes that do not actively counter this tendency reproduce it. The second underrepresentation is occupational. The populations most exposed to consequential AI deployments — healthcare recipients in public systems, beneficiaries of social protection programs, smallholder farmers subject to algorithmic credit scoring, workers subject to algorithmic management — are rarely present in governance processes. Their experience constitutes irreplaceable diagnostic data about where governance frameworks fail in practice. The third underrepresentation is linguistic-cognitive. Global AI governance discourse is conducted almost exclusively in English and structured around legal-technical frameworks derived from Anglo-American regulatory traditions. This excludes not only non-English speakers but entire theoretical traditions — Brazilian sociotechnology, African Ubuntu ethics, Andean Buen Vivir frameworks, Islamic legal epistemology — that offer genuinely distinct governance architectures for managing the relationship between algorithmic systems and human communities. Inclusion requires more than translation. It requires redesigning the deliberative process so that non-Northern normative frameworks can be presented as governance proposals, not merely as cultural context.

Three engagement formats would meaningfully differentiate the AI Dialogue from the accumulated record of multilateral governance declarations that have produced normative consensus without operational consequence. Governance failure audits as deliberative anchor. Rather than beginning from abstract principles, each thematic session should be anchored in a structured comparative analysis of documented AI governance failures — cases where AI systems deployed in public infrastructure produced measurable harm, with evidence about which governance mechanisms were absent or insufficient. Comparative analysis across multiple jurisdictions, presented by the researchers and civil society actors who documented the failures, produces a shared diagnostic that frames subsequent deliberation around concrete accountability questions rather than aspirational language. This format forces the question: what would have prevented this specific failure? — a question that existing frameworks have conspicuously avoided answering. Normative prototyping workshops. A distinct track should bring together technical researchers, legal experts, and community representatives to prototype governance mechanisms — not principles, but operational specifications: what would a computationally verifiable human oversight requirement look like? What audit schema would allow a government in a developing country to certify that an AI system deployed in its health sector cannot optimize away constitutional rights? These workshops should produce draft technical annexes, not position papers, and their outputs should feed directly into the Dialogue's formal deliberation track. Affected community testimony with institutional response obligations. Testimony from communities directly harmed by AI system failures — patients denied care by opaque triage algorithms, farmers excluded from credit by biased scoring models, workers subject to algorithmic discipline — should be structured not as witness statements but as governance stress-tests: each testimony should be formally responded to by delegations from relevant jurisdictions, specifying which governance mechanism would have prevented the documented harm and what commitments are made toward closing the gap. This format converts testimony into accountability, not merely into acknowledgment.

Three concrete approaches merit attention as exemplars of effective AI governance, selected because they represent distinct points on the spectrum from institutional design to technical architecture. Brazil's SUS constitutional framework as a normative accountability model. Brazil's Unified Health System embeds universality, integrality, and equity as justiciable constitutional rights - not aspirational principles. This means that any AI system deployed in public health resource allocation is, in legal terms, subject to judicial enforcement if its outputs systematically violate these mandates. The framework demonstrates that grounding AI governance in constitutional rather than merely regulatory norms produces a qualitatively different accountability structure: one where affected populations have standing to challenge AI-mediated decisions through existing legal channels without requiring new AI-specific legislation. The weakness of this model - that constitutional norms are expressed in natural language requiring judicial interpretation - points directly to the next frontier: mechanisms capable of translating constitutional mandates into computationally verifiable constraints. The OECD AI Principles sectoral working groups, particularly AIHEG in health, demonstrate that domain-specific expert governance can produce comparative evidence on governance gaps that generic risk frameworks cannot generate. The sectoral approach - developing governance criteria calibrated to the specific normative context of healthcare, food security, or education - is more likely to produce actionable standards than horizontal frameworks that treat all high-risk AI as structurally equivalent. Neurosymbolic audit architecture as an emerging technical practice. Research demonstrating that formal logic-based normative constraints can be embedded directly in AI system architectures - making non-compliance structurally impossible for defined classes of violations rather than externally detected after deployment - represents the most consequential frontier in AI governance engineering. Pilot deployments in public infrastructure that produce empirically validated comparisons between compliance-by-verification and compliance-by-construction would constitute the most significant contribution the current generation of governance research could offer to the Dialogue.