National Forensic Sciences University,India

The first Global Dialogue on AI Governance will be meaningful only if it moves beyond declarations and produces something tangible that governments and communities can actually use. From my perspective, success would look like this: First, agreement on a shared baseline vocabulary. Right now, the word "AI governance" means entirely different things in Brussels, Nairobi, Jakarta, and Washington. Without terminological alignment, every subsequent conversation starts from scratch. The Dialogue should produce a working glossary that is genuinely negotiated, not imported wholesale from one regulatory tradition. Second, a clear mapping of who is doing what globally. Dozens of frameworks, national strategies, and multilateral initiatives already exist. The Dialogue should honestly catalogue these — including their contradictions — rather than pretending each is complementary. Third, a firm commitment to structured follow-through. A one-time meeting with no accountability mechanism serves mainly as a photo opportunity. Success means leaving Geneva with assigned working tracks, named focal points, and a reporting expectation before the 2027 New York session. Fourth, genuine evidence that underrepresented stakeholders shaped the agenda — not just attended. If the final summary reads as though it was written before the room assembled, the exercise has failed. Fifth, at least one concrete, implementable deliverable directed at countries with limited AI governance infrastructure. Whether that is a modular policy toolkit, a peer-learning network, or a technical assistance referral mechanism, something practical must emerge that smaller and developing nations can carry home and use within twelve months. Ambition is appropriate. Vagueness is not.

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

These four areas are not arbitrary choices. They reflect where the gaps between rhetoric and reality are widest, and where inaction carries the most direct human cost. Safe, secure and trustworthy AI is the foundation. Without it, every other governance discussion is decorative. We are already seeing AI systems embedded in criminal sentencing, loan approvals, medical triage, and border control - often without meaningful safety evaluation. The risks are not theoretical. AI capacity-building is urgent for a specific reason that is often stated gently but deserves to be said plainly: the majority of the world's countries are currently recipients of AI governance norms, not authors of them. That asymmetry is itself a governance failure. Building genuine technical and institutional capacity in the Global South, in small island states, in least-developed countries - this is not charity. It is the condition for any governance framework being legitimate and sustainable. Protection and promotion of human rights must be explicit, not implied. There is a persistent tendency in technical governance discussions to treat human rights as a value to be "balanced" against innovation or economic growth. That framing is wrong. Rights are not a variable. Surveillance systems, predictive policing tools, and content moderation algorithms are already causing documented harm. The Dialogue must name this directly. Transparency, accountability, and human oversight closes the loop. Governance without enforceability is aspiration. Transparency requirements - including meaningful audit rights, disclosure of training data provenance, and algorithmic impact assessments - give other principles actual teeth. Human oversight is not anti-innovation. It is what makes trust possible. These four together form a coherent logic: trustworthiness as the goal, capacity as the precondition, rights as the boundary, and accountability as the mechanism.

Yes. Three issues deserve explicit attention that the current thematic list either omits or addresses only indirectly. Environmental impact of AI systems. The computational infrastructure powering large AI models - data centres, cooling systems, energy draw - has significant and growing environmental consequences. This is not a niche concern. It intersects directly with climate commitments and with energy equity in countries where AI infrastructure competes with basic power needs. Any governance framework that ignores this is incomplete. AI and labour displacement. The economic implications of AI are listed in the thematic areas, but labour displacement deserves its own sustained focus rather than being folded into a broader category. The transition risks are unevenly distributed - clerical, administrative, and routine service workers in developing economies face acute exposure without the social protection systems that some wealthier countries have in place. This needs governance attention now, not after the disruption has occurred. Governance of frontier and general-purpose AI systems. Current frameworks are largely built around specific applications. General-purpose AI systems - those capable of performing across many domains - do not fit neatly into sectoral regulation. The Dialogue should explicitly address how international governance applies to these systems, including questions of liability, evaluation standards, and which actors bear responsibility when harm occurs across jurisdictions. Concentration of AI power. A handful of private firms, concentrated in two or three countries, currently control most of the foundational AI infrastructure the world depends on. This market concentration is itself a governance issue with implications for sovereignty, competition, and the distribution of AI benefits. It should be named, not navigated around.

Working across multiple regions in a research and advisory capacity, the patterns I observe are consistent enough to describe with confidence. In lower-income countries, the primary challenge is not absence of will — it is absence of infrastructure. Many governments want to develop AI governance frameworks but lack trained regulatory staff, legal precedents adapted to their contexts, or access to technical expertise that is not tied to a commercial vendor with a product to sell. They are being asked to govern systems they did not build, using frameworks designed elsewhere, without the resources to evaluate whether those frameworks fit their circumstances. In the research and academic sector globally, the governance gap manifests as a mismatch between the pace of AI deployment and the pace of independent evaluation. By the time rigorous third-party assessment of a system is published, that system has often already been widely adopted. There is no international mechanism that creates a structured role for independent researchers in pre-deployment review of high-risk AI systems. This needs to change. The opportunity that the Dialogue represents is precisely the convening power of the UN — its ability to bring together actors who would not otherwise sit in the same room and to legitimate frameworks that have no single powerful backer. That convening power is real and should be used deliberately, not squandered on process for its own sake. The most significant risk I see is that the Dialogue produces a document that satisfies everybody at the level of language while committing nobody to anything measurable. That outcome would be worse than silence because it would consume political energy and trust that could otherwise support something real.

The AI Dialogue has a realistic opportunity to do something no existing forum has managed: create a genuinely inclusive space where AI governance is negotiated rather than announced. Existing mechanisms — the EU AI Act, the OECD AI Principles, the G7 Hiroshima Process, the US Executive Order framework — are valuable, but they share a common limitation: they were written by a small number of actors and then offered to the rest of the world as a starting point. The Dialogue, operating within the UN system, has a different mandate and a different legitimacy base. Concretely, the Dialogue can advance cooperation in three ways. Norm translation. Not every country needs an identical regulatory framework, but there are core principles — around risk classification, human rights obligations, and transparency — that can travel across different legal systems if they are translated properly. The Dialogue can resource that translation work systematically. Early warning and information sharing. When an AI system causes harm in one jurisdiction, that information rarely reaches regulators in other countries in a structured way. The Dialogue could catalyse a lightweight international information-sharing mechanism — not a full regulatory body, but a structured channel — that allows harm documentation and incident reporting to cross borders. Mutual recognition and interoperability. Rather than creating yet another governance standard, the Dialogue could focus on establishing conditions under which different national frameworks can recognise each other's evaluations. This reduces duplication and lowers the compliance burden on actors operating across multiple jurisdictions, while allowing countries to maintain regulatory sovereignty. The Dialogue should resist the temptation to try to do everything. Focused ambition is more productive than comprehensive ambition with no follow-through.

Several existing initiatives offer genuine building blocks, and the Dialogue would be stronger for connecting with them honestly rather than treating them as competitors. The OECD AI Policy Observatory has done substantial work cataloguing national AI strategies and identifying policy trends. Its methodology is rigorous and its data is publicly accessible. The Dialogue should not duplicate this — it should draw on it and push it toward broader geographic coverage. UNESCO's Recommendation on the Ethics of AI is the only globally negotiated instrument specifically on AI ethics that has been adopted by a broad, diverse set of member states. Its implementation support programme deserves more resources and more visibility. The Dialogue should treat it as a foundational reference rather than one document among many. The ITU's AI for Good platform has developed practical tools and convened technical communities that the Dialogue's processes would benefit from. The AI governance conversation must not be separated from technical reality. Regional frameworks — the African Union's AI strategy, ASEAN's AI governance framework, and the emerging discussions within CELAC — represent governance thinking that emerged from different development contexts. The Dialogue should actively draw these into the global conversation rather than treating them as junior partners to OECD-derived frameworks. The Global Digital Compact, adopted in September 2024, established principles and commitments that the Dialogue should be explicitly aligned with. Coherence between UN processes matters. The added value the Dialogue can bring is not another document. It is the convening of actors who are not currently in conversation with each other, and a structured commitment to follow-through that none of these initiatives individually can provide.

The format of the Dialogue will determine whether its outputs are legitimate. Procedure is not bureaucratic detail here — it is substance. On structure: The two-day Geneva meeting should not be organised primarily around plenary speeches. Plenary sessions should be reserved for synthesis and decision points. The substantive work should happen in smaller, thematically focused working groups where genuine exchange is possible and where written outputs are produced in the room, not drafted by secretariats afterward. On stakeholder roles: Each stakeholder category brings something distinct and irreplaceable. Governments bring legal authority and implementation capacity. Private sector actors bring technical detail and operational reality. Civil society brings accountability pressure and the voices of affected communities. Academia brings independent analysis. The technical community brings implementation knowledge. None of these can substitute for another, and none should dominate. On practical inclusion: Remote participation must be genuine, not performative. This means time zone-sensitive scheduling of substantive sessions, real-time interpretation in multiple languages beyond the six official UN languages where feasible, and advance circulation of working documents with enough lead time for organisations without large research teams to prepare meaningful inputs. On outputs: Each working group should produce a structured summary — not a consensus communiqué that sands down every disagreement, but an honest account of where convergence exists, where it does not, and why. Disagreement documented clearly is more useful than false consensus. On continuity: The Dialogue needs a light-touch permanent secretariat function between sessions — not a new institution, but a designated coordination point that maintains relationships, tracks commitments, and prepares the 2027 session based on what the 2026 session produced.

The underrepresentation in global AI governance discussions is systematic, not accidental, and addressing it requires deliberate structural choices rather than open-door invitations. Affected communities — people who have experienced harm from AI systems in domains such as social benefit administration, criminal justice, border control, and content moderation — are almost entirely absent from governance conversations. Their expertise is experiential and direct, and it is precisely the kind of expertise that prevents governance frameworks from becoming self-referential. Civil society organisations from the Global South frequently lack the resources to participate in Geneva-based processes even when they are formally invited. Travel funding, translation support, advance access to documents, and meaningful agenda influence — not token panel slots — are the concrete things that inclusion requires. Indigenous communities have articulated specific concerns about AI systems that affect their languages, their cultural heritage data, and their land rights. These concerns are almost entirely absent from mainstream AI governance texts. Small and medium-sized enterprises from developing economies are subject to AI governance frameworks but rarely consulted in their design. Their operational realities differ substantially from those of large technology corporations. Workers and labour organisations are the people most immediately affected by AI-driven changes to employment, yet their representation in AI governance forums is minimal compared to employer associations and technology companies. Practical inclusion mechanisms include: dedicated funding for participation from underrepresented regions; thematic sessions specifically designed around the experiences of affected communities; partnership with regional civil society networks to aggregate and represent voices that cannot individually attend; and a commitment that final documents be reviewed for whether they meaningfully reflect input from underrepresented groups before adoption.

Innovation in format should serve substance, not substitute for it. With that caveat, several approaches could genuinely improve the quality of dialogue. Structured red team exercises. Rather than presenting governance proposals for comment, participants could be organised into groups tasked with identifying the most significant failure modes of a proposed approach. This surfaces problems that polite panel discussion tends to avoid. Reverse testimony sessions. Instead of experts presenting to affected communities, affected community representatives present to policymakers and technical experts, who are required to respond substantively. This shifts the power dynamic in the room in ways that improve the quality of governance thinking. Real-time gap mapping. Using collaborative digital tools, participants could simultaneously populate a shared document identifying governance gaps in their jurisdictions during the session itself. Aggregated in real time, this produces a living map of where the most urgent needs are concentrated — a more useful output than another list of principles. Cross-regional peer review. Pairs or small groups of countries with different development levels and regulatory contexts could review each other's draft governance approaches and provide structured feedback during the session. This builds relationships and produces practical learning simultaneously. Asynchronous deliberation before the event. The most substantive governance discussions should not begin on the first morning in Geneva. A structured pre-dialogue process — thematic working papers circulated in advance, comment periods, synthesis documents — would mean that the in-person time is used for genuine deliberation rather than initial position-stating. Youth-led scenario sessions. Structured scenarios about AI governance challenges in 2030 and 2035, facilitated by youth delegates, would bring both a longer time horizon and a different generational perspective into the conversation in a substantive rather than symbolic way.

Several concrete examples from different contexts offer practical lessons. Canada's Directive on Automated Decision-Making is worth studying carefully. It requires federal departments to assess the impact level of any automated decision system before deployment, disclose its use to affected individuals, and provide human review options. It is not perfect, but it is operational, it covers a broad range of government applications, and it has produced measurable changes in how departments procure and deploy AI tools. The methodology of tiered impact assessment adapted to different contexts is transferable. Singapore's Model AI Governance Framework is notable for being sector-flexible and for providing implementation guides alongside the framework itself. Governance frameworks that consist entirely of principles without implementation guidance create compliance theatre rather than genuine accountability. Singapore's approach of connecting the principle to the practice is a model worth adapting. Rwanda's National AI Policy demonstrates that countries with limited current AI infrastructure can develop thoughtful, contextually grounded frameworks rather than simply importing models from elsewhere. Its explicit attention to using AI for domestic development priorities while managing risks reflects a governance posture that other developing nations could learn from. The EU's AI Act, despite its limitations and the ongoing challenges of implementation, has established that legally binding risk-based regulation of AI is achievable. Its classification methodology - general prohibitions, high-risk requirements, and transparency obligations - provides a structural vocabulary that other jurisdictions can engage with even where they do not adopt it wholesale. Brazil's AI Bill, still in legislative process, represents an important case of a large developing economy designing its own framework from the ground up rather than harmonising with frameworks designed for different economic and social contexts. The process itself - including multi-stakeholder consultation over several years - is as instructive as the output. The common thread across these examples is that effective governance is specific, operational, and honestly implemented. General commitments to "responsible AI" produce nothing. Specific requirements with clear compliance mechanisms produce change.