UNDP

Success at Geneva in July 2026 should not be measured by the ambition of declarations but by the specificity of commitments. Three outcomes would mark genuine progress. First, agreement on a tiered accountability framework — distinguishing governance obligations for AI systems that make high-stakes decisions affecting individuals and organizations (financial gatekeeping, benefits determination, law enforcement) from those with lower consequence profiles. The current tendency toward uniform principles obscures the urgency of targeted intervention where harm is already occurring. Second, a formal mandate for intersectoral working groups that include not only Member States and technology companies, but practitioners from multilateral institutions, civil society, and Global South regulatory bodies. The AI governance debate has been shaped disproportionately by OECD-based actors. An inclusive Dialogue that does not structurally correct for that asymmetry will produce standards that reflect it. Third, a concrete capacity-building commitment with a funding mechanism — not a pledge, but a dedicated instrument, modeled on the GEF or the Technology Facilitation Mechanism, that allows low- and middle-income countries to participate in AI governance as rule-makers rather than rule-takers. Technical sovereignty requires resources, not only rhetoric.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights
• AI capacity-building

These four areas are not independent priorities - they form an integrated accountability chain, and their selection reflects where AI governance failures are already producing measurable harm. Safe, secure and trustworthy AI is the foundational commitment, but its practical meaning depends on who defines "trustworthy" and in whose context. AI compliance tools used in sanctions screening, AML/CFT, and development finance due diligence are marketed as trustworthy while systematically producing biased outputs against non-Western name structures, fragile-state entities, and civil society organizations operating in conflict-affected areas. Safety must be operationalized, not aspirational. Transparency, accountability, and human oversight are the mechanisms without which safety commitments are unenforceable. AI systems making high-stakes decisions - blocking payments, flagging counterparties, denying access to finance - must carry explainability obligations and meaningful human review requirements. Voluntary disclosure is insufficient where the affected parties lack leverage to demand it. Protection and promotion of human rights grounds the governance framework in existing international legal obligations. The right to remedy, the right to non-discrimination, and due process protections apply to AI-driven decisions just as they apply to human ones. The AI Dialogue should affirm this explicitly rather than treating human rights as a soft complement to technical standards. AI capacity-building is the equity condition for everything else. Without it, low- and middle-income countries will remain subject to AI governance standards they had no role in designing, deployed in tools they cannot audit, producing outcomes they cannot contest. Capacity-building must mean technical sovereignty - the ability to develop, evaluate, and regulate AI - not only training in how to use systems designed elsewhere.

Two structural gaps deserve explicit recognition. AI in financial gatekeeping systems. The use of AI in sanctions screening, AML/CFT transaction monitoring, counterparty risk scoring, and development finance due diligence constitutes one of the most consequential deployments of automated decision-making in the international system - yet it appears in no thematic cluster. These systems determine who accesses the global financial system, which organizations receive multilateral funding, and which jurisdictions face correspondent banking withdrawal. They are predominantly built by a small number of private vendors in OECD countries, trained on data that underrepresents the Global South, and deployed with limited transparency or redress mechanisms. The AI Dialogue should establish a dedicated workstream on AI in financial gatekeeping, with participation from FATF, FSB, UNODC, and affected civil society. AI and virtual asset-enabled illicit finance. The intersection of AI and crypto-assets represents one of the fastest-evolving governance gaps in the international system. AI is being weaponized by illicit actors - through synthetic identity generation, automated chain-hopping, and mixer obfuscation - at a pace that exceeds the detection capacity of most national authorities. FATF Recommendation 15 on virtual assets is not yet calibrated to this dual-use dynamic. The AI Dialogue should engage FATF and the Egmont Group to develop governance standards specific to AI-powered illicit finance through virtual assets, and should ensure that capacity-building on crypto forensics reaches Global South jurisdictions as a structural equity commitment, not an afterthought. Both issues share a common structural feature: the actors who bear the consequences of AI deployment in these domains - fragile states, civil society organizations, diaspora communities, humanitarian implementers - are least represented in the rooms where governance standards are set. The Dialogue's inclusive mandate is only meaningful if it actively corrects for that absence.

From the perspective of an AML/CFT practitioner with experience across multilateral institutions and international banking in Latin America, Sub-Saharan Africa, and fragile-state contexts, the governance gaps in the four priority areas produce concrete, measurable harm in three interconnected ways. Financial exclusion through algorithmic de-risking. AI-powered risk scoring has made it operationally cheap for correspondent banks to exit entire corridors. Latin American and Caribbean remittance corridors, Palestinian civil society organizations, and humanitarian implementers in conflict-affected states have all experienced accelerated de-risking directly linked to automated compliance decisions. The opacity of these systems means affected entities cannot identify the basis for exclusion, let alone contest it. This is a human rights issue with a technical face. Asymmetric standards in multilateral due diligence. UN agencies and multilateral development banks apply AI-assisted screening tools to counterparties across 170+ country offices — often using commercial systems not designed for development finance contexts. The result is disproportionate scrutiny of local civil society partners, community-based organizations, and government entities in fragile states, while the tools themselves are audited by no international body. There is no governance framework that applies to the AI systems of multilateral institutions equivalent to what those same institutions demand of their private sector counterparts. Capacity asymmetry in regulatory oversight. Most Global South financial intelligence units and regulators lack the technical capacity to audit AI compliance tools deployed within their own jurisdictions by international banks and fintech operators. This means AI governance in practice — the actual decisions that affect real people — is occurring in a regulatory vacuum in the jurisdictions most exposed to harm. The opportunity is equally concrete: a well-designed AI governance framework for these sectors could reduce false positives, restore access to finance for legitimate actors, and give regulators in low-capacity jurisdictions enforceable standards to invoke.

The AI Dialogue's most distinctive contribution to international cooperation is not standard-setting — other bodies do that — but legitimacy construction. It is the only AI governance forum with universal membership, a UN mandate, and an explicit inclusion commitment. That combination creates something no OECD club, industry consortium, or regional body can replicate: a space where small island states, fragile economies, and civil society practitioners can engage as co-authors of governance frameworks rather than recipients of them. To realize that potential, the Dialogue should play three specific cooperative roles. Forum of forums. The AI governance landscape is fragmented across OECD.AI, GPAI, FATF, ITU, UNESCO, regional bodies, and bilateral frameworks. The Dialogue cannot — and should not — replace these. But it can serve as the connective architecture that identifies gaps between them, resolves conflicts between overlapping standards, and ensures that the Global South has a meaningful voice in each. A formal mapping and coordination function would give the Dialogue structural relevance beyond its convening sessions. Norm translation mechanism. High-level AI principles — transparency, accountability, human oversight — mean different things in different regulatory and developmental contexts. The Dialogue can facilitate the translation of agreed norms into sector-specific guidance, with working groups that include practitioners alongside diplomats. Financial integrity, health, and education each require contextually grounded governance — not uniform application of abstract principles. Early warning and review function. The Dialogue should establish a mechanism for identifying AI governance failures as they emerge — not only at the 2027 session, but on an ongoing basis. A lightweight intersessional review process, with civil society and expert input, would allow the Dialogue to remain responsive to a rapidly evolving landscape between formal convening cycles. Cooperation without accountability architecture is aspiration. The Dialogue should build both.

Several existing mechanisms provide foundation the Dialogue should explicitly engage rather than duplicate. FATF and the FSB set the de facto global standards for AI in financial compliance. The Dialogue should establish a formal liaison with both bodies — not to replicate their technical work, but to ensure that the human rights and equity dimensions of AI-powered AML/CFT are incorporated into their methodology reviews. FATF's 2025–2030 strategic priorities include digitalization and AI; the Dialogue has a window to influence that agenda before it closes. UNESCO's Recommendation on the Ethics of AI (2021) represents the broadest intergovernmental consensus on AI principles to date. The Dialogue should treat it as a baseline, not a competing framework — building sector-specific operationalization on top of agreed ethical foundations rather than relitigating first principles. The ITU's AI for Good platform and the OECD.AI Policy Observatory provide technical infrastructure and comparative data the Dialogue lacks. Rather than building parallel systems, the Dialogue should formalize data-sharing and joint publication arrangements with both. The UNODC and Egmont Group are the primary international bodies addressing AI in illicit finance and virtual assets. The Dialogue should commission a joint study with both on AI-enabled financial crime typologies and governance gaps — creating an evidence base that neither body has produced alone. The Technology Facilitation Mechanism (TFM) under the 2030 Agenda already provides an intergovernmental structure for technology governance. The Dialogue should operate in explicit coordination with the TFM to avoid duplication and leverage existing multi-stakeholder engagement infrastructure. The added value the Dialogue brings to all of these is universality and political legitimacy. None of the above bodies has universal membership or a General Assembly mandate. The Dialogue can give their technical work the intergovernmental authority needed to produce binding or quasi-binding governance outcomes — if it chooses to use that mandate seriously.

Meaningful multi-stakeholder participation requires structural design, not open-door rhetoric. The Dialogue's Geneva session should distinguish between three participation modes — each with distinct functions, not merely different seats at the same table. Deliberative participation — reserved for Member States and formally accredited stakeholder groups — should produce the session's substantive outputs: prioritized governance gaps, draft principles, and mandates for intersessional work. This requires small enough working groups to enable genuine exchange, not plenary statements read into the record. Expert input — from practitioners, academics, civil society, and technical bodies — should be structured as evidence sessions preceding deliberation, not side events appended to it. The distinction matters: expert input that informs deliberation changes outcomes; expert input that decorates it does not. Each thematic workstream should have a mandatory expert briefing, with written submissions made publicly available before Member States convene. Affected community testimony — from individuals and organizations directly impacted by AI governance failures — should be formally integrated into the program. Not as a symbolic gesture, but as evidentiary input that shapes the framing of governance gaps. Communities experiencing AI-driven financial exclusion, surveillance, or labor displacement have knowledge that neither governments nor technical experts possess. On structure: the Dialogue should move away from the standard UN conference format — general debate followed by negotiated text — toward a problem-led architecture. Each session should open with a clearly defined governance failure, draw on practitioner and affected-community evidence, and task working groups with producing concrete recommendations rather than consensus language that papers over disagreement. Intersessional continuity is equally important. The 2026–2027 gap between Geneva and New York should be filled by thematic working groups with defined mandates, civil society participation, and public reporting — not silence between formal sessions.

The governance gap in AI is also a participation gap. The communities most exposed to AI's harms are least present in the rooms where its governance is designed. Four groups require structural, not merely aspirational, inclusion. Financial integrity practitioners in the Global South. Financial intelligence units, AML/CFT compliance officers, and regulators in low- and middle-income countries operate daily within AI-powered systems they did not design, cannot audit, and have no formal channel to critique internationally. Their practitioner knowledge of how these systems fail — in which corridors, against which entity types, with what consequences — is irreplaceable governance intelligence currently absent from every major AI forum. Civil society organizations operating in conflict-affected and fragile contexts. These organizations are among the most systematically harmed by AI-driven compliance systems — flagged, de-banked, and excluded from development finance through automated processes with no redress. They are also among the least resourced to engage in Geneva-based UN processes. The Dialogue should establish a dedicated civil society fund covering participation costs, with priority for organizations from fragile and conflict-affected states. Indigenous communities and linguistic minorities. AI systems trained on dominant language datasets perform poorly — and sometimes harmfully — in indigenous language contexts, affecting everything from content moderation to public service delivery. These communities are structurally absent from AI governance processes. Regional pre-consultations in indigenous languages, with outputs formally transmitted to the Dialogue, would begin to correct this. Young people outside OECD contexts. Youth engagement in AI governance has been disproportionately channeled through Northern-based organizations and entrepreneurship frameworks. The Dialogue should partner with regional youth networks — particularly in Africa, South and Southeast Asia, and Latin America — to integrate perspectives grounded in the actual AI environments young people in these regions inhabit. Inclusion requires resource allocation. Without it, the Dialogue's multi-stakeholder mandate is a procedural fiction.

The risk for the AI Dialogue is that it defaults to the UN conference format: prepared statements, negotiated bracketed text, and a communiqué that satisfies no one. Geneva in July 2026 should be designed differently — not as novelty, but because the subject demands it. Structured adversarial debate. For each thematic workstream, assign practitioners and Member State representatives to argue opposing governance positions — not their own — before open deliberation. This surfaces genuine disagreement rather than allowing it to be smoothed into consensus language, and produces more durable agreements because the tensions have been named. Governance failure case studies as session anchors. Each thematic session should open with a documented case — a de-risked remittance corridor, an AI-generated false positive that blocked humanitarian funding, a deepfake-enabled financial fraud — presented by those directly affected. Abstract principle debates are transformed when grounded in specific, verifiable harms. This is standard practice in human rights bodies; the AI Dialogue should adopt it. Real-time drafting with public visibility. Working group outputs should be drafted in real time on publicly visible platforms, with a defined window for civil society comment before finalization. This is not radical transparency — it is the minimum condition for legitimacy in a multi-stakeholder process that claims to be inclusive. Intersessional digital participation. Between Geneva and New York, the Dialogue should operate thematic digital forums — not webinars, but structured asynchronous deliberation platforms with moderated expert facilitation — allowing participation from jurisdictions that cannot sustain travel costs across two UN conference cities in consecutive years. Written contributions from these forums should carry formal weight in the 2027 session. Regional pre-consultations with transmitted outputs. Before Geneva, regional bodies — AU, ASEAN, CELAC, SICA — should host structured AI governance consultations whose outputs are formally transmitted and acknowledged in the Geneva program. Inclusion that begins at Geneva is already too late.

Several existing approaches offer concrete models the Dialogue should study and build upon - not as templates for uniform adoption, but as evidence that specific governance designs produce better outcomes. The EU AI Act's risk-tiered framework is the most developed attempt to calibrate governance obligations to consequence severity. Its prohibition of certain AI applications, mandatory requirements for high-risk systems, and lighter-touch approach to lower-risk uses provides a structural model - even for jurisdictions that would design the tiers differently. The Dialogue should examine its implementation gaps and the equity concerns raised by its extraterritorial reach on Global South AI developers and deployers. FATF's Guidance on AI in AML/CFT (2023) demonstrates that sector-specific AI governance - not only horizontal principles - is both feasible and necessary. Its framework for explainability, human oversight, and model risk management in financial compliance contexts should be extended, updated to address virtual assets, and given stronger implementation benchmarks. The African Union's AI Continental Strategy represents an important counter-model: a governance framework developed from a sovereignty and development lens rather than a safety-and-competition lens. The Dialogue should treat it as equal in authority to OECD and EU frameworks, not as a regional complement to Northern standards. UNODC's GlobE Network on corruption and financial crime demonstrates that practitioner peer-learning networks - rather than only treaty-based mechanisms - can produce operational governance improvements at speed. An equivalent network for AI governance practitioners in financial integrity, health, and public administration would generate the evidence base that formal intergovernmental processes currently lack. Rwanda's AI Policy and Uruguay's algorithmic transparency decree illustrate that small and middle-income countries can produce governance innovation - not only import it. The Dialogue should systematically document and elevate these cases, making the Global South a source of governance models, not only a recipient of them.