Harvard Carr and Ryan Center for Human Rights

Success would mean the Dialogue resists the temptation to paper over the fundamental tension at the heart of AI governance: we are trying to build global rules in an era of deliberate decoupling. The two most consequential AI powers are not converging but rather they are building parallel ecosystems, parallel standards, and parallel visions of what AI is for. Any outcome that pretends this is merely a technical coordination problem will not survive contact with reality. A successful Dialogue names this honestly and asks the harder question: what governance is actually possible under these conditions, and for whom? For countries caught between these competing systems and there are many, including in my own region the absence of a credible multilateral framework is not an abstraction. It means adopting standards designed elsewhere, for interests that are not yours, with no seat at the table where those standards were set. Success would mean the Dialogue creates genuine entry points for these countries, not as recipients of AI governance but as co-authors of it. Concretely, I would consider it a success if the Dialogue produced one durable institutional mechanism not a declaration, not a roadmap, but an actual body with a mandate to review AI's human rights implications on an ongoing basis. The UN human rights architecture exists. The question is whether member states have the political will to use it for AI. Ambition is easy. Institutional memory is hard. That is what I would measure success

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights

These four priorities reflect what I see as the irreducible minimum for AI governance that is both effective and legitimate. Safety and trustworthiness without accountability is marketing. Accountability without human oversight is procedure without power. I selected these together because they are only meaningful in combination, a system can be technically "safe" by narrow metrics while still producing outcomes that violate rights or concentrate power without recourse. The social, cultural, and linguistic implications of AI are systematically underweighted in governance conversations dominated by technical and legal expertise. For smaller nations and linguistic communities including my own AI systems trained on dominant language data encode assumptions that do not translate. This is not a minor localization problem; it is a structural equity issue. Human rights provides the only existing international framework with both normative weight and implementation mechanisms. Anchoring AI governance there is not idealistic it is practical. It gives states, civil society, and affected communities a common language and existing institutional pathways to raise concerns and demand remedy.

The most significant gap is geopolitical fragmentation as a governance variable. The listed themes implicitly assume that global cooperation is the default condition. It is not. The bifurcation of AI development between major powers, with divergent standards, export controls, and competing infrastructure ecosystems creates a structural problem that smaller and middle-income countries navigate daily, with no multilateral framework to support them. A second gap is AI's role in information environments and electoral integrity, particularly in contexts of active foreign interference. This sits uneasily across several listed themes but is captured by none fully.

I work at the intersection of AI governance and human rights, with a focus on the South Caucasus and post-Soviet space, a region that rarely appears in mainstream AI governance conversations, and almost never as an actor rather than a subject. Georgia is a striking case. It is a small hybrid democracy under sustained pressure from democratic backsliding domestically and from Russian information operations externally. AI-enabled disinformation is not a future risk there; it is a present condition. Yet Georgia has no AI strategy, no dedicated regulatory body, and no meaningful seat in the forums where norms are being set. The governance gap is not abstract it shows up in the daily information environment that shapes political outcomes. More broadly, the South Caucasus sits at a geopolitical fault line between competing AI ecosystems. Countries in the region face pressure to adopt infrastructure, platforms, and standards designed in Moscow, Brussels, or Washington and each carrying embedded assumptions about data sovereignty, state access, and acceptable use. There is no multilateral framework that helps smaller states navigate these choices on their own terms. In my sector, AI governance research and policy , the gap I encounter most is institutional: the absence of review mechanisms that can assess AI's human rights implications in real time, in specific contexts, with standing to demand accountability. The UN human rights architecture exists and has legitimacy. What is missing is the political will to extend its mandate clearly into AI. That is what I am working on. And that is why this Dialogue matters not as a one-time event, but as a potential foundation for something more durable

The most honest answer is: a limited but non-trivial one, if it is designed with that limitation in mind. We are not at a moment when the major AI powers will agree on binding rules. That window may not open for years. What the Dialogue can do is hold space for everyone else, the countries that are neither setting the agenda nor building the frontier systems, but who will live with the consequences of both. For those countries, the Dialogue's value is not in producing a grand treaty. It is in creating a legitimate forum where smaller states can compare notes, build coalitions, and develop shared positions before they are presented with fait accompli standards designed elsewhere. That kind of structured solidarity is underrated and genuinely hard to build without a UN-level convening mechanism. The Dialogue can also do something quieter but important: establish a shared language. Right now, "AI safety" means very different things in Geneva, Washington, and Beijing. Alignment on definitions , even partial, even contested is prerequisite to any deeper cooperation.

The Dialogue should resist the urge to start from scratch. The UN human rights treaty bodies, the Special Procedures system, and the work of the High Commissioner's office on technology and rights already provide a normative foundation that has taken decades to build. The added value of the Dialogue is not replacing this, it is connecting AI governance explicitly and operationally to it. Beyond that, the OECD AI Principles and the Council of Europe's AI Convention represent real attempts at binding or semi-binding frameworks. The Dialogue should acknowledge what they got right rather than produce a competing document. What it can add that these cannot: genuine universality. The OECD is a club of wealthy democracies. The Council of Europe is regional. The Dialogue is the first forum where the Global South, small island states, and conflict-affected countries sit at the same table as the frontier powers. That is rare. It should be used.

The Dialogue's format will determine its politics. If it runs like a standard UN conference, national delegations, prepared statements, negotiated language, it will produce negotiated language and nothing else. The stakeholders who most need to contribute are not the ones who already know how to work a UN room. Civil society organizations from conflict-affected states, researchers working in non-dominant languages, indigenous communities whose data is being harvested without consent , these groups have direct knowledge of AI's impacts that no government delegation can replicate. But they need structured access, not just observer status. That means dedicated working sessions, written submission pathways that are actually read, and rapporteurs who are accountable for incorporating their input. On format: I would strongly recommend thematic tracks that run in parallel and feed into plenary, rather than a single sequential agenda. AI governance is too broad for one conversation. Smaller, expert-driven tracks produce better outcomes and allow genuine dialogue rather than statement-reading

The most underrepresented voices are not hard to identify, they are simply inconvenient to include properly. Small and lower-middle-income countries in the Caucasus, Central Asia, and Sub-Saharan Africa experience AI's effects daily in their information environments, their border systems, their labor markets, but contribute almost nothing to the standards governing those systems. This is partly resource-driven: AI governance participation requires legal expertise, technical fluency, and travel budgets that many governments simply do not have. The fix is not symbolic. It requires funded fellowships for practitioners from underrepresented regions, translation of core documents beyond the six UN languages, and a standing civil society advisory mechanism with real access, not a side event. I would also flag linguistic communities specifically. AI systems trained predominantly on English, Mandarin, and a handful of European languages encode cultural assumptions invisibly. Georgian, Armenian, Kazakh, and dozens of other languages spoken by millions are effectively ungoverned territory in AI development. Their speakers have no meaningful recourse when systems fail them.

The most valuable format innovation would be a structured red-teaming process, not for AI systems, but for proposed governance frameworks themselves. Before any text is adopted, it should be stress-tested by practitioners from regions most likely to be harmed by its gaps or loopholes. This is standard in good policy design and almost never done in multilateral settings. Beyond that: asynchronous participation mechanisms matter more than they are given credit for. Not everyone can be in the room. A well-designed digital platform for structured input, not open comment boxes, but guided, thematic, time-bound submissions would dramatically expand who can meaningfully engage. Finally, the Dialogue should build in a feedback loop: a commitment to report back to participants within 12 months on which inputs were incorporated and why others were not. Accountability for the process itself is what makes future participation worthwhile.

The most instructive examples are not always the most celebrated ones. The EU AI Act gets significant attention, and rightly so it is the first comprehensive attempt to regulate AI by risk category with binding obligations. But its real contribution is less the specific rules than the underlying logic: that AI governance should be proportionate, that high-risk applications require human oversight, and that rights-based impact assessment should be mandatory rather than voluntary. That logic is exportable even where the specific legal architecture is not. Less discussed but worth highlighting: the Council of Europe's Framework Convention on AI is the first internationally binding treaty on the subject, and crucially, it is open to non-member states. For countries outside the EU and Council of Europe including in my region, this is significant. It creates a pathway to alignment with rights-based standards without requiring membership in institutions that may not be accessible. At the practice level, I am struck by the work being done by civil society organizations in conflict-affected and transitional contexts, documenting AI-enabled disinformation, building community-level digital literacy, creating accountability records when formal institutions cannot. This work rarely appears in policy papers, but it is often the only governance that actually functions on the ground. What these examples share is that they treat affected communities as sources of knowledge, not just recipients of protection. That inversion from subjects to co-authors of governance is the most important shift the Dialogue could encourage and the hardest to institutionalize.