UN Transcripts — https://transcripts.un.org/en/asset/k17/k17pcqctx1 Thematic discussion 3: Safe, secure and trustworthy AI - Global Dialogue on Artificial Intelligence Governance - Day 2 — 7 July 2026 Language: en Transcripts available through this tool are created by using automatic speech recognition and are not official records nor official documents of the United Nations. Official records and official documents are available on the Official Document System of the United Nations. --- Speaker 1 [19:17]: Good morning, good morning, colleagues. We request you all to please take your seats. We're now about to begin this session. Just a quick reminder, please do check if your name is on the list of inscribed interventions available on the Dialogue website, and if it is, please do make your way to the reserved speaker row on the right-hand side of the auditorium, the second row. For all speakers who will be up on the panelists, please do take your seats in any of the front rows. With this, we are very pleased to hand over to your facilitators for today. We have UNICC and UNDRR, over to you. WIPO [20:06]: Thank you so much, Excellencies, distinguished delegates, colleagues. Good morning and welcome to Thematic Cluster 3 on Safe, Secure, and Trustworthy AI: Interoperability and Compatibility of Approaches. Around the world, we use different electric electrical systems. Different plug shapes, different voltages, yet we travel, our laptops and phones still work safely. Not because every country uses the same system, but because we have developed adapters, common standards, and trusted interfaces that allow different systems to connect. AI governance faces much the same challenge. As AI systems increasingly operate across borders, countries will continue to develop different laws, different regulatory traditions, and different policy priorities. Harmonization is neither realistic nor entirely necessary. But interoperability is essential. Without it, governance frameworks become fragmented. Implementation becomes more complex and costly. Trust is harder to build and the benefits of AI become more difficult to realize and to share. At WIPO, helping different national legislation and systems work together is not a new challenge. It is what we have been doing for decades. And in the age of AI, WIPO's AI Infrastructure Interchange Initiatives is working to build the interoperability needed for the future. Interoperability is not about creating a single governance model or lowering standards. It's about enabling different approaches to communicate and to work together, grounded in shared evidence, common understanding, and practical cooperation. Today's discussion in Cluster 3 brings together governments, international organizations, industry, academia, and civil society to explore how we can build those foundations together. It is now my pleasure to hand over to our thematic cluster co-chairs. First of all, Her Excellency Paula Bragantes Zamora. You have the floor. Costa Rica · Co-Chair · Paula Bragantes Zamora [22:21]: Thank you. Buenos días. Good morning, everyone. Excellencies, distinguished colleagues. The world does not need more AI principles. It needs a common way to prove they're being implemented. That is the real interoperability challenge before us, not whether frameworks can coexist on paper, but whether safety, accountability, and trust can be demonstrated across borders. This challenge is unfolding within an ecosystem of extraordinary concentration. In 2025, institutions in the United States produced 59 notable AI models, and China produced 35. While the rest of the world produced only 13. The United States also held roughly 75% of the computing power among the 500 largest known AI clusters, with China holding another, another 15%. Concentration of infrastructure becomes concentration of evidence. The countries and companies with the greatest computing capacity also have greater influence over, over which risks are measured, which benchmarks are accepted, which languages are evaluated, and which institutions are recognized as capable of certainty— certified safety. This is why the lack of participation matters. The preliminary scientific assessment indicates that 118 countries, mainly from, from the Global South, are not meaningfully engaged in the principal international discussions of AI governance, while fewer than one-third of developing countries have national AI strategies. We should therefore resist two false choices. Interoperability must not mean regulatory uniformity, particularly when regions begin from profound different technological foundations. In 2025, 5G networks cover 70% of the population in Asia Pacific, but only 12% in Africa. In 2024, Africa attracted just 3% of global data center investment. A governance architecture that assumes universal access to advanced connectivity, computing resources, and specialized regulatory expertise would reproduce inequality rather than overcome fragmentation. At the same time, respect for national context must not become an excuse for systems that cannot communicate, compare evidence, or recognize one another's safeguards. The objective must be minimum practical compatibility across terminology, risk classification, data protection, cybersecurity, documentation, testing, and incident reporting, while allowing countries to sequence implementation according to their institution— institutional capacity and preserve legitimate legal and development priorities. Language is another decisive test. The world has more than 7,000 languages, but AI training and evaluation cover only a fraction. Models often perform less safely in languages with limited digital data. From our region, that means Spanish variants, Portuguese, indigenous, and Creole languages cannot remain outside global evaluation frameworks. This challenge becomes especially concrete when AI systems operate across institutions and jurisdictions. A health system, judicial tool, or public sector algorithm cannot be responsibly transferred, evaluated, or scaled if countries use incompatible definitions of risks, documentations, requirements, or audit results. The challenge becomes even more urgent with agentic AI, where systems may act on behalf of people or institutions across multiple platforms. In that environment, interoperability must also include verifiable agent identity, traceable delegation, machine readable permissions, secure data exchange, audible actions, and effective revocation. Estonia Aarau— Aarauite Initiative illustrates this emerging frontier. It is developing identity and registry frameworks, interoperability standards, legal recommendations, and a public sector pilot so that state can interact with AI agents in a trustworthy and accountable way. The lesson is clear. Future interoperability must connect not only systems but rules, but also identities, mandates, and chains of responsibility. Costa Rica proposes a practical starting point, Minimal Viable Interoperability. Before 2027 dialogue, we should advance shared terminology, comparable risks classifications, consistent documentation, interoperability comparable incident reporting and multilingual evaluation methods while building pathways towards mutual recognition of testing, auditing, and certification. Interoperability will succeed when evidence can travel, accountability can be traced, safety claims can be verified, and every country has the capacity to help define the rules. That is the foundation on trust of trustworthy AI, and that is the work before us today. Lastly, I would like to recall a powerful point made by my fellow co-chair Rebecca Finlay. Adoption moves at the speed of trust. That insight goes to the heart of today's discussion. Trust cannot be declared through principles alone. It must be built through evidence, safeguards, institutional cooperation, and governance frameworks. That can operate across borders. Rebecca, the floor is yours. PAI · Co-Chair · Rebecca Finlay [28:36]: Thank you. Thank you so much. Thank you for those opening words, and thank you to the co-chairs, to the organizers, and everyone in this room for all you are all doing to move us forward today. Two moments stood out to me yesterday. Secretary-General Guterres stressed the importance of common baselines for frontier systems, common methods to verify them, and resolve to ensure systems meet common standards for global trust. Maria Ressa grounded her address in real-world examples, positioning scientific evidence as the antidote to both fearmongering and techno-utopianism. The panel, in her words, scrutinized the most contested findings and demanded the most rigorous evidence. I heard a clear message that I agree with: the panel and the dialogue are two parts of a shared mission. The panel provides the evidence, the dialogue provides the direction. Evidence and deliberation. Designed to work together from the outset. So here we are. What are we going to do together this morning? I'd like to recommend 3 concrete actions to come out of today and to move forward over the next year. First and foremost, strengthen the independent evidence base. Secondly, open up the science behind it. And third, turn evidence into progress that the public around the world can see and hold us to. First, strengthen the independent scientific evidence-based. Invest in and connect the work of the scientific panel. Drive coherence by connecting to other state of safety reports, whether it's from the UK, the recent Singapore consensus, or the forthcoming Global South Safety Report, link to the International Network of AI Safety and Security Institutes, deepen evaluation science globally. Get the dialogue right and it becomes the connective tissue binding these efforts into one shared foundation for the world. Second, prioritize and open up the science. Transparency doesn't slow progress. It's what makes progress verifiable, credible, and worth defending. Strengthen disclosure from private companies who are developing and deploying AI globally. Build a coherent international assurance ecosystem. Disclosure alone isn't enough. Without that system to make sense of it— benchmarks, standards, third-party eval— evaluation and verification. Let's turn individual disclosures into a field-wide evidence base that can be trusted. Third, we must advance progress in the public interest. A safety standard that is set by a few is not accountable and it is not safe. Define a shared baseline of good practice inclusively, and let's track progress against it. Use baselines such as existing human rights frameworks, and let's measure ourselves against them. Build on the commitments that reflect real consensus today on what good looks like. For the most severe risks and the most vulnerable groups, such as children, we must take precautionary action based on the advice of independent experts over time. With 170 countries here today and a rich mix of industry, civil society, academia, and government, each of the UN dialogues starting today can measure where the field and the fight actually stands against that baseline and assess what's still missing. At the Partnership on Aids, AI, we believe that no single country, government, nonprofit, nor research institution can divine on its own what safe, secure, and trustworthy AI looks like. This week, we announced the Global AI Progress Hub, a public platform where organizations across the responsible AI ecosystem can share their progress against a very clear framework that serves the public interest. Closing the gap between innovation and accountability starts with asking different questions than the ones that are currently driving the field. Instead of, "How fast can we deploy this?" let's ask, "Who does it serve? Who is accountable when it fails, and how do we know?" Instead of, "How powerful is this system?" Let's ask, how verifiable is this system safety, and against what evidence? AI systems must be built inclusively, governed accountably, and operated fairly, transparently, reliably, and securely. The work of getting there is why this dialogue exists. I'm looking forward to the conversation this morning. Thank you. UNICC · Facilitator [34:20]: Thank you very much to our co-chairs for those opening remarks. On behalf of all of us here and on behalf of UNICC, we'd like to now move on to a scene-setting presentation which will help frame the discussions that will follow throughout this morning. And it is my distinguished pleasure to welcome Under-Secretary-General and Special Envoy for Digital and Emerging Technologies, Amandeep Singh Gill. Amandeep, the floor is yours. Thank you very much. Microphone for the speaker. UN · USG and Special Envoy for Digital and Emerging Technologies · Amandeep Singh Gill [35:05]: And if our friends are now— I think the mic is working. Good morning, everyone. Welcome to the second day of the dialogue, and thank you, co-chairs Paula and Rebecca, for setting the scene for this morning's very important discussion. And it's good to see the distinguished co-chair Ambassador Tham-Sar here in the room. So this cluster is very important because it links three important themes. The objective of building safety and trust— yesterday, the Secretary-General, when he spoke about what four things, the four things that we must get right, safety was on top of his list. And those questions must be linked to the practical question of how different governance approaches can work across borders and the role that this dialogue could play in fostering interoperability horizontally. But let's first remind ourselves of the challenge. Yesterday, when the panel presented— the scientific panel presented its report, it emphasized the shift from static to dynamic system governance. And our work through the AI Governance for Humanity Lab in Valencia has also turned up evidence of this shift from static model oversight to dynamic system governance, especially as agentic AI systems become more autonomous, tool-using, and multi-step. This rapid evolution can outpace existing governance practices which are still being adapted to systems whose behavior may change over time. Effective governance will need to be more adaptive and anticipatory, with mechanisms such as living risk taxonomies, controlled testing environments, and staged deployment to identify risks before systems are widely ruled rolled out. This of course raises the bar somewhat for our discussion on interoperability, and let me spend a few moments on this issue. And to better understand it, perhaps we can think about what's the opposite of interoperability, the antonym of interoperability., and that's fragmentation. So global fragmentation creates several structural challenges. First, it leads to regulatory arbitrage, as developers and vendors may move activities to jurisdictions with weaker oversight, weakening safety, trust, and human rights protections across borders. Second, it creates accountability gaps. Since AI systems operate through global value chains, this makes it difficult to determine applicable law or assign responsibility when harms occur. Third, fragmented rules increase compliance burdens, especially for SMEs, researchers, and developers from low-income countries. Who may lack the resources to navigate multiple regimes. And finally, fragmentation deepens uneven governance capacity and dependency, as lower-capacity states may be forced either to align with dominant frameworks that do not reflect local realities, or risk exclusion from global data flows, trade, and and investment. Now obviously, if we get it right, we have less fragmentation, greater interoperability, we can foster more inclusive AI innovation ecosystems and promote innovation across these global value chains. Now, How can we operationalize interoperability? So here I'm sharing some pathways for reinforcing interoperability. These are pathways, called so deliberately because these are not prescriptive recommendations or a fixed roadmap, but practical guideposts for cooperation. And I'll just give one example of a— one of these pathways to illustrate the potential utility of this approach. So cross-border regulatory sandboxes, which connect together regulatory sandboxes from different environments and can be used to test AI applications under regulatory supervision before they are deployed widely. This pathway proposes establishing connections between regulatory sandboxes to enable cross-border test pilots, model evaluations, and evidence sharing. So, what I'm getting at is that the objective is not to harmonize AI governance into a single model, but to build practical bridges across diverse approaches so that AI can be safe, secure, and trustworthy across borders. And this brings me to my last point, which is the role of this dialogue, the role of the United Nations as the most inclusive platform for discussions on AI governance. This is not a top-down platform. It is in fact a horizontal facilitation of safe, secure, trusted, and interoperable AI governance. This is a new approach to international learning. If we look at the model in many other domains, it's essentially the, the multilateral model is from the 1950s, and that works very well in certain domains, but AI is a different beast. So we need a new approach to promoting international learning across borders, across stakeholders, and that's the important and essential work that we've started yesterday with the inaugural Global Dialogue on AI Governance, informed by the work of the independent International Scientific Panel. Thank you very much. UNICC · Facilitator [42:07]: Thank you so much to Ahmadinejad for setting the scene so effectively. We will now transition to our first moderated panel. To guide us through the discussion, I'm pleased to welcome our moderator, Miss Virginia Dicknam, Professor of Umeå University, on stage. She will introduce our distinguished panelists and lead the conversation. Miss Dicknam, please come up and call up your panelists The floor is yours. Umeå University · Professor · Virginia Dignum [42:56]: Good morning, and thank you all. Thank you for inviting me to moderate this panel. AI is not an inevitability that we must simply adapt to. It is a set of choices—technical, institutional, and political. Operability is one of the clearest examples of this type of choice. Without shared evidentiary standards and mutually recognized approaches, we risk parallel regimens that neither protect people nor allow accountability to travel with the technology. We have seen this before in many different fields where fragmentation creates gaps that harm the least resourced first. And joining me on discussing this today, I have Saeed Ahmed from Infosys, Melat Bilce-Dirmikos from AI Panel, Nouf Al Ameli, a special advisor to the UAE, Yoshi Ashida, government from the government of Japan, and Leonard Cervera Navas from the European Data Protection Supervision, the Director General there. So please, I would like to welcome you all to the stage. This side here, please. Yeah. I'll go— We have 25 minutes only. And I will be very strict on the 3 minutes that are allocated for each of your presentations. And I would like to start with the first question. And this question will be addressed to Mr. Yoshihide Iida from the Government of Japan and to Syed Ahmed from Infosys. The question is, which existing international legal frameworks, legal and policy instruments technical standards, and other initiatives could be used as already existing blocks for establishing common foundation and fostering interoperability in AI governance? Mr. Yoshii, I will give you 3 minutes, and I'm counting. Japan · Advisor and Visiting Professor · Yoichi Iida [45:34]: Okay, thank you very much, moderator, and good morning, co-chairs, and good morning to everybody in the Conference room. My name is Yoichi Iida, advisor at the ministry and as well as the visiting professor at Tokyo University. So let me pick up one of our very important instrument, which are OECD AI principles and Hiroshima AI process. As many of you know, OECD AI principles are the set of very high-level value-based principles. And we believe this may be one of the foundational instruments for interoperability. In fact, Japan formulated our domestic guidelines and our national AI law based on these principles. And also, we discussed our international initiative called Hiroshima AI Process Hype. Based on these principles together with our colleagues from G7 group. So Hiroshima Process provide a code of conduct for AI organizations, including AI developers, and code of conduct provide a voluntary framework for those organizations to assess and mitigate risks and also to disclose relevant information on their risk management. So these frameworks are the very ambitious multi-stakeholder approach for collaborative governance. And as Amandeep stated, this is a bottom-up approach for the governance. So actually, we have already started our work for interoperability with colleagues from ASEAN countries, in particular Singapore. So we are now working on the cross-mapping and crosswork between HYBE framework and the ASEAN framework for advanced AI models. And we found some of the gaps, but we found a much bigger number of commonalities between these two frameworks. And in most cases where we found gaps, Many of them can be explained by the difference in the diversity in social or cultural aspects from individual jurisdictions. So I think similar work can be applied to different combinations of national or regional frameworks, and I hope we can contribute from our experience. Thank you very much. Speaker 10 [48:46]: Thank you very much. Thank you. And I would like now to give the word to Ahmed— Sayed Ahmed. Thank you very much. Microphone for the speaker, please. Infosys · Sayed Ahmed [48:55]: Thank you very much. I represent the GSI community. I work for Infosys. One of the reasons why this impacts us the most is because we operate globally and the compliance burden on adhering to each and every regulations and standard becomes much higher. Now coming to your question, it doesn't matter where we start. We can definitely start like how my colleague here mentioned with OECD or ISO. ISO has this 22982 that can be enhanced, or EU, or even what we have, the AI standard exchange databases. But what is important is we cover the entire length and breadth of all the regulations and standards that we have. When we start building the adapters, we need to build such that it covers everything, it cannot leave out, and it's quite complicated. The reason is, in one way, we talk about AI regulations, the multiple AI regulations across the geographies and things like that. Then you have tech regulations, it needs to be interoperable with other tech regulations like say GDPR and DORA and a lot of other regulations that may be there. Then again, the geography comes in. This puts a lot of compliance burden on company like ours, which operates globally. And just a shared taxonomy is not going to cut it, or it is not going to be adapter. We tried being an AI company, we tried solving this using AI. We thought it's very easy. We'll build an AI model to build an adapter where it can map it. It was not easy. One thing that we can learn from our failure is when you start mapping, I'll give you an example. Say, for example, transparency. Overall, in the intent, it'll mean same thing, say, in the EU Act. And say, for example, in ISO, in ISO, transparency means you need to be explainable. Your model need to be explainable. In EU Act, it means that you need to be able to provide sufficient evidences. You need to have the logs, the system logs and all that. It is not one-to-one mapping, even if the intent largely is the same. So the complication is when we are Building these adapters, you need to go 2, 3 steps down and to be able to map each of the controls that are required to be covered in this. So we need to go down and just say, yeah, based mapping, it's quite easy, right? You can go to any model and say, these are the 3 standards, get me a mapping of it. It'll give you a high level, the first job. The thing that we need to invest on and do as a committee here is build that meta model that is required for the controls mapping at the lowest level and build that knowledge graph that can then use as a basis for— that should be the foundation stone so that we can build on top of it for our interoperability models. Thank you very much. Moderator · Virginia Dignam [51:41]: And now I'd like to move to the second question, which is, where is fragmentation or lack of a common evidence— base already creating practical challenges for safety, for trust, accountability, innovation, or the implementation of existing governance approaches? Which stakeholders, sectors, regions are bearing the greatest costs of that fragmentation? And I would like to start with Ms. Nouf Al Ameli. Science and Technology Advisor · Nouf Al Ameli [52:13]: Thank you. And good morning everyone. My name is Nofel Hamli. I'm the Science and Technology Advisor in the Ministry of Foreign Affairs. And I guess I'm going to address this question more from a cross-border rather than a cross-sectorial perspective. So one place that I would point to is the fact that high-risk AI doesn't mean the same thing to any two regulators right now, right? So we have to treat this challenge as a linguistic a legal problem as well as a technical problem. You can see this with how the EU AI Act defines risk tiers largely by use case. You can see this by the way that the US approach, for example, and where it exists at least, it has a large sectorial and voluntary definition for it. And then there are some other countries that require algorithm registration and content labeling that don't map onto each other, and that probably many other examples in state-based domestic policies. So a single model deployed globally can be high risk, it could be lightly regulated, it requires state filing simultaneously depending on which border it crosses. It's the operating reality for any company that's being deployed abroad right now, and I think that's the main focus on our fragment— fragmentation at the moment. Right now, the governance frameworks are multiplying faster than the evidence base that's underneath them. And there's no shared incident reporting infrastructure. So you'll have different incident monitors tracking failures independently and voluntarily, but there's no legal duty to report or to recognize, uh, or any recognition between them. So if a safety failure occurs in one country, there is no guaranteed path for that lesson to reach regulators elsewhere before the same failure recurs somewhere else. Um, one angle that I don't think gets enough airtime, and this is the part that's more— that isn't focused in cross— on cross-border, but rather in the cross-sectorial within the countries themselves. And this is something we feel we face in the UAE as well. So a state might regulate AI and banking rigorously because it already regulates banking. The same country may have zero equivalent oversight over other areas, but because it doesn't— but it's not because it doesn't recognize those as high risk, but because there's no pre-existing regulatory owner or owning of that domain. Domain. And that domestic sectoral fragmentation is arguably— it's much more fixable. It's immediate than the international one. It's a useful one. It's a useful thing to name before pivoting harder to the harder ones, which is cross-border cases. As an example, I wanted to bring in an example that's much more closer to home. So in the UAE, I come from a research background. And in the research area, it's always been very difficult to align research governance research governance across government, industry, and academia within the triple helix model. This has been hard even with three actors that have shared national interest. So that's a small-scale preview. If you want to apply this to more than 190 countries, that makes the challenge a lot more strong— a lot more harder. We felt this enough domestically, and as of June of this year, the UAE folded three mandates that were overseen by different federal entities our AI office, the digital government arm of TDRA, which is the Telecommunication Regulatory Authority, and the Emirates Data Office. Moderator · Virginia Dignam [55:32]: Can you please wrap up? Science and Technology Advisor · Nouf Al Ameli [55:33]: Sure. So the reason we did all of this is we wanted to show that consolidation is not coherent— that it can stay coherent. So I'll end this with saying that the answer isn't conversions for its own sake. We can push too hard on one thing. But we can have— but if you have one standard, you could risk a lowest lowest common denominator outcome that can cost us the innovation system as well and might have a risk on the national interests in those sectors. I'll stop there. Moderator · Virginia Dignam [56:01]: Thank you very much. Now I would like to give the word to Melat Bilciu-Demircos. Melat Bilciu-Demircos [56:07]: Hello. It's a pleasure to be here after 3 months of hard work from the UN AI Panel, and I'm very happy that our report is out. So I'm going to base my answer today on the report that's been distributed. So our report identifies four areas where the structural challenge is very evident. So the first one is evidence dilemma. So what we are seeing is that robust scientific evidence cannot catch up with how fast AI is progressing because it doubles, the capacity doubles every few months, and by the time we have evidence, real harm or its capacity is compiled and technology has already passed beyond current governance models. And we've found 40— we've documented 40 different governance models and ethical guidelines. And what we find, as you— as has been mentioned, is they're highly fragmented, inconsistent, and rarely tested. And if they are tested, they're tested by the same companies that are developing them. So this results in systems that evade clear accountability, and especially as has been mentioned again across borders. But in my case, so this, in this case it's my field, we also find that there are a lot of ambiguities when legal frameworks are applied also out to space, not just to Earth. So this introduces transboundary conditions not just on Earth, out to space. Now, one thing that we wanted to emphasize in our report was that when we are looking at our— at systems, we shouldn't just be looking at the models, but we should have a more approach that looks at the interoperability not just of models, but also of tools of the environment of users. So we should shift from evaluating just models to evaluating systems. And because we lack a unified independent evidence that is outside of the companies that are developing these models, this undermines public trust and objective accountability of these companies. The final one, the fourth one, is unclear microscopic and innovation outcomes. We haven't found consistent complete data on how data, AI is affecting economic outcomes. So there is a critical gap, especially in common evidence regarding labor and productivity. We can see that it is narrowing task efficiencies, but lucky, especially long-term macroeconomic growth and/or clear labor market displacement trajectories are lacking right now. And this makes proactive evidence-based economic policy also very difficult. Now, coming to your question about stakeholders— Moderator · Virginia Dignam [59:05]: You have 5 seconds. Maybe you can leave that for closing remarks and we move to the next question. Melat Bilciu-Demircos [59:15]: Sorry. Thank you. Moderator · Virginia Dignam [59:16]: We have to be very strict like we had indicated before. I would like to move then to the final question to Mr.— I lost the names— Leonardo Cervera Navas, Director General at the Data Protection Authority of the European Union. And the question is, what can be done to ensure that interoperability is built into AI governance from the start? What happens if interoperability and the development of technical standards and other solutions to assess interoperability are treated as an afterthought. The floor is yours, 3 minutes again. EDPS · EU · Director General · Leonardo Cervera Navas [59:59]: Excellencies, distinguished colleagues, it is a real honor for me to participate in this UN event representing the European Union. I'm more in particular representing the authority in charge of supervision of AI in the EU institutions and bodies. I congratulate the United Nations for this initiative because the unprecedented concentration of wealth and economic power that we see in the development of AI needs to be balanced with a similar concentration of state power and moral purpose and responsibility, because technology must serve mankind and not the other way around. We are here to discuss governance of AI, but the real question is governance for what? For us, As Europeans, we got the answer in the legal basis of the AI Act. In Article 114 of the Treaty on the Functioning of the European Union, we say that we need governance for the protection of health and safety. And in Article 16 of the same treaty, we say that we need governance for the protection of privacy of individuals and other fundamental rights. But the starting point should never be technophobic. The EU embraces this technology and its potential for advancing science and general well-being. And the fact that it comes with some clear risks does not mean that we should oppose AI. I'd like to mention the example of the first transformative technology of humankind, controlled fire. Our ancestors learned to control this technology so they could live in the caves and keep the predators away and keep warm during the night and to cook, and that made them truly humans. But they quickly realized that this technology of fire came with risks. So they developed two simple rules to apply to this technology. You always need to use fire with care, and you can never leave fire not supervised. So exactly the same rules should apply to AI. We have to be careful with this technology and we always have to have this technology under supervision. This is exactly what we have done in the EU with our AI Act. We have established principle-based rules where we say that the— yeah, the higher the risk, the higher the care and the supervisions. So I think I stop here and I will continue a little bit later. Moderator · Virginia Dignam [1:03:17]: Thank you. Thank you very much. We are almost at the end, so I would like to give all the speakers 1 minute to give their concluding remarks. And it's really 1 minute. I will move to the next one after 1 minute. I start with Mr. Ashihira. Japan · Advisor and Visiting Professor · Yoichi Iida [1:03:34]: Okay, thank you very much. So I think interoperability is very important. That is because, you know, we have to respect diversity as well. So thinking about the fragmentation, we need to understand how we can achieve and promote interoperability through the bottom-up and also multi-stakeholder approach. And that is So that's what we have been doing and we hope we can work together with a wider group of countries and other stakeholders in this framework of United Nations. So thank you very much and we are very much excited to work together, all together, and we look forward to further work. Thank you. Moderator · Virginia Dignam [1:04:27]: And we move to Noura. Just start. Noura [1:04:30]: All right, I'm going to try and make this very quickly, very quick 3 points. So I think one thing is that that full compatibility in our interoperability, I think it's been mentioned a couple of times here, that that approach doesn't need to be the goal. And I think we need to really work on selective and deliberate compatibility of the components that matter the most for most nations. And that might be a good place to start. The last thing I'm going to share is probably the 3 principles that the UAE anchors itself. And when we're thinking about our approaches, given that we change our governance structure, especially when it comes to AI, so that it represents the ongoing or sort of changing technological pace that we see in front of us. The first principle is on the trust between the partners of the nations, which is a precondition for any shared framework to hold under constant change and pressure, which really applies here with interoperability, whether it's cross-border or cross-sectorial within the country. The second one is credibility, which is ensured by the accountability both within the governments and between them and having commitments that are verifiable and not just declared in statements. And then the last and probably the most important is sovereignty and respecting that the states have different starting points and different paths. And I think just using those as guiding principles when we're thinking about interoperability will allow us to start from a good place. Thank you very much. Moderator · Virginia Dignam [1:05:51]: Birgit? Birgit [1:05:53]: So I'll talk about the stakeholders. You'd asked about— in your question, you'd also asked about Which stakeholders, sectors, regions bear the greatest cost? What we find in our report is that developing nations face the steepest cost of this fragmentation as frontier models are based and compute is concentrated in just two countries around the world. Now, the Global South therefore risks being locked out completely because they don't have localized infrastructure. And a common— and a common global evidence base which they can use. So these regions are reliant on technologies which they lack domestic capabilities of, and they— so they cannot inspect, audit, or culturally tailor these models to their own needs. And clearly there are vulnerable populations such as children and women. The rise in AI-generated child sexual abuse as well as deepfake-enabled sexual violence has increased. And we find that civil society and democratic frameworks are also being damaged. There is an erosion of shared reality as well as an undermining of the democratic legitimacy due to automated disinformation and AI-generated content. Moderator · Virginia Dignam [1:07:12]: Thank you very much. Birgit [1:07:13]: Sorry, my last sentence, please. Moderator · Virginia Dignam [1:07:15]: Okay. Birgit [1:07:16]: And end users. I think one thing that our Secretary-General stated yesterday, but I find it worth repeating, is mental health. That the lack of rigorous safety baselines are— and caused by the sycophantic chatbot behaviors are affecting mental health around the world. Thank you. Moderator · Virginia Dignam [1:07:41]: Very important. Thank you so much. Leonardo. Thank you. EDPS · EU · Director General · Leonardo Cervera Navas [1:07:46]: I mentioned before the analogy of fire. I would like to now mention the analogy of civil aviation. Is there anything more risky and dangerous than putting people in planes, taking up to 10,000 feet and move them around the world? It is a super risky activity, but yet It works extremely well. Why? Because we have the same rules applicable everywhere in the world. No matter if you take a plane in Kabul or if you take a plane in New York, the same safety rules apply everywhere. This is exactly what we need to make AI a success for business and for citizens. Let's agree on some basic safety and ethical rules and have them applied all around the world. Thank you so much. Thank you very much. Moderator · Virginia Dignam [1:08:45]: Thank you. Infosys · Sayed Ahmed [1:08:47]: And that's very good. I'll try my best to cover my closing remarks in 60 seconds. Now, I would like to answer the question that you had asked earlier about do we need to have AI governance AI governance as from the start of developing the AI use cases, not as an afterthought. I kind of disagree on that because we need to separate AI governance from the AI use cases building. The reason is what Undersecretary General also did mention, saying that AI governance has now become more dynamic. It cannot be static. We can't start at the start of the use case and assume the same rules will hold for 10 years down the line or even 1 month down the line. Because things are changing so fast. So what we need is separation of AI governance versus the AI use cases and ability to be able to apply policy as a code for any use cases separately. So we should be able to apply— build certain central repository of policies that can be applied in the runtime and enforced in the runtime so that we can separate this effectively. This is my thought. Thank you. Moderator · Virginia Dignam [1:09:53]: Thank you very much. And I realize it's awful for all of you that I have to cut you all, but that was my— what was ordered of me. And I would like to thank you all and thank all the participants for this very short but very important discussion. Thank you all. Thank you very much. Facilitator [1:10:28]: Hello. We have an updated list. Okay. Because things have been moving very quickly. But Rivka also has a list to share. Okay. And I think maybe I'd like to thank Professor Dignam and the distinguished panelists for a very interactive session. We'll now hand the floor back to our co-chairs who will manage the floor for interventions from member states and from stakeholders. Over to you. Moderator · Virginia Dignam [1:11:08]: Good morning, distinguished delegates, excellencies, colleagues. It's wonderful to see so many people here for this discussion. We're now going to facilitate the interventions in the segment before moving on to the second panel. So, statements will be delivered on the basis of the speakers' list established through the inscription that was made available on the UN Global Dialogue on AI website. So, in the interest in promoting broad participation from stakeholders, we'll proceed with one representative from member States, followed by one representative from other stakeholders, alternatively. So if you have been— if you have inscribed and you've received the message that you will be speaking today, I ask you to please take your seats in the front two rows so that you can come up to the stage to speak from the lectern. The e-delegate list is closed, and any changes has to be communicated to the secretariat in the room. You'll find Filippo and Antonio over there. Thank you. A timer has been set and will be visible on the screens as well as on the lectern. We kindly encourage all speakers to speak in the allotted time so that we can facilitate the largest number of speakers possible. Microphones will automatically be switched off once your time has expired. With that, I hand over to the co-chairs. Thank you. Co-Chair [1:12:35]: We begin with Austria, followed by UPU and Latvia. Okay, so we will begin with UPU, followed by Latvia and EY. UPU [1:13:39]: I just want to check the mic's working. Thank you very much. Thank you very much, co-chairs and excellencies, colleagues. I speak on behalf of the Universal Postal Union, which is a United Nations specialized agency for the postal sector. It links 192 member states in a single cross-border network for the exchange of international mail. Um, the panel's preliminary report offers this dialogue a sobering common starting point. Current safeguards are not keeping pace with AI capabilities,, and most states are depending on systems that they can neither inspect nor audit. Governance instruments remain fragmented and are rarely tested against real-world effectiveness. The report further warns that new risks can emerge from the interaction of multiple autonomous agents acting across connected systems. From the perspective of the Universal Postal Union, and in particular the postal supply chain, these are not abstract concerns. One stakeholder's automated decision, a customs flag, a routine choice, a risk score immediately affects others down the chain. What the report describes as a systemic risk is for us an everyday operational reality. And in that context, ladies and gentlemen, The UPU underlines two priorities for safe, secure, and trustworthy AI, and we've heard some of this today. The first is interoperability. Without a shared governance baseline, AI in our sector might develop in proprietary and uneven ways, deepening the very divides the panel describes and leaving smaller and developing country operators behind. Interoperability of governance approaches which is the theme of this cluster, cannot, however, stop at the level of principle. And this was a point picked up in the panel previously, this idea that principles such as transparency actually require further deep dive to ensure that we have a common level of understanding as it— as how it might actually apply in a particular sector. And this is something that we are quite focused on. The second is accountability and oversight. High-impact automated decisions, for instance, that hold, at least in the postal supply chain, a shipment, clear it, or determine how it is treated, must remain under clear human accountability and be traceable and contestable. The panel cautions that there is no scientific guarantee that autonomous agents will follow their instructions and that evidence of them departing is already accumulating. Accountability and oversight must therefore be built by design and not reconstructed from failure. The UPUs response, ladies and gentlemen, has been over the few months to convene a voluntary— sorry, I think I've lost the mic again. Co-Chair [1:16:56]: Thank you very much. Now we have the distinguished representative from Latvia who will be followed by EY Ernst Young and then Brazil. Latvia [1:17:12]: Dear colleagues, thank you, distinguished co-chairs. Latvia thanks Estonia, El Salvador, and the Joint Secretary-General for preparing this dialogue. We highly value its truly multistakeholder nature. Today our societies face a paradox. AI increasingly delivers profound benefits. It has predicted the structures of proteins to design new cures. In Latvia, it is detecting stroke with 93% of accuracy. Yet reported incidents of serious harm is rising and our citizens react. Despite the benefits, trust in AI decreases, not increases. Globally, 7 in 10 call for stricter regulation. So this is a strong mandate for us to ensure that AI can be trusted. But trust must be demonstrated through practical verifiability, transparent governance, data policies, and independent testing of high-risk applications throughout their lifecycle. And it must be complemented by innovation-supporting mechanisms such as regulatory sandboxes, just mentioned, to validate high-risk AI systems before full-scale deployment. Supporting innovators rather than punishing it. That's the approach we have taken in Latvia. Therefore, Latvia strongly supports risk-based AI governance grounded in a shared, verifiable, evidence-based, safe, secure, and trustworthy— AI is not only a technical challenge, it is also a matter of information integrity, cultural and linguistic diversity, and public interest. Here, Latvia experience offers two lessons. First, on information integrity. While gathering data for an open EU multilingual AI model, we had to filter out millions of articles from coordinated disinformation network infiltrating global AI models. Poisoned training data silently distorts essential facts and downstream the worldview built on them. Yet there is no shared standard for training data origin and no channel to warn others. Second, on underrepresented languages. Latvian language is spoken by 2 million people. Evidence confirms that gap between dominant and underrepresented languages in leading AI models are widening, not narrowing. This is unequal. It strips away cultural heritage and context, but it also costs several times more tokens when using AI in small languages. Yet the gap can be closed with multi-stakeholder partnership and development of open models. As an example, EU-supported OpenWaits till the open model achieved equality across more than 30 EU languages. So Latvia proposes 3 priorities: shared evaluation methods and open benchmarks reflecting linguistic and cultural diversity, minimum compatibility on data origin and incident reporting including data integrity incidents, and capacity building so every country can participate in the evaluation as an equal partner. Thank you. IAHWG · Co-Chair [1:20:09]: Thank you, Latvia. Now we have Ernst Young, followed by Brazil and Sage Bionetworks. Okay, so now we're followed by Brazil. Ernst Young is here, sorry. Brazil · Eugênio Garcia [1:20:36]: Thank you, Chair. Eugênio Garcia from Brazil. Co-Chair [1:20:41]: Apologies for the delay. Sorry, Brazil. We're going to have now Ernst Young and then followed by Brazil. Thank you. EY [1:20:48]: Thank you. Apologies for the delay. Ernst Young, as a global firm that is operating in virtually every country in the world, supporting businesses in their adoption of AI systems, but also importantly supporting businesses in, uh, their auditing of whether or not systems are operating the way that they should do, is seeing a very strong demand for interoperability between the different governance approaches that should be taken in different member states of the UN. Obviously Different countries have different ways of operating. They have different, um, uh, different political frameworks. Uh, so they— we, we cannot expect all countries to approach, uh, the regulation of AI systems in exactly the same way. But interoperability does not require this. Interoperability can live at the implementation stage to a large extent. So this is obviously an area where technology standards play an important role that can be referred to in the regulatory approaches. But it is also a question around making sure that our approach to making— to good governance of AI is rooted in the core question of making sure that AI operates in a way that functionally delivers what is needed. And this is where I think the business and industry community can really meet the same demands and, and has the same kind of questions as society as a whole does. Nobody wants the AI systems to not perform in the way that they were meant to do. This is bad for society. It's also bad for the industry that is trying to use these tools. So really emphasizing that Good governance is not about compliance. Good governance is about making sure that the AI systems operate in an appropriate way, can be a core basis for interoperability in our approaches to doing AI governance. Thank you very much. Co-Chair [1:23:00]: Thank you very much. Now we will have the distinguished representative now from Brazil and then Sage Bionetworks. Brazil · Eugênio Garcia [1:23:07]: Networks and France. Thank you, Chair. Eugene Garcia from Brazil. Trusting AI systems cannot be separated from inclusion, transparency, accountability, human oversight, and respect for human rights, democracy, and the rule of law. Safe and trustworthy AI must not be understood only as a technical issue. It is a political, social, and developmental challenge. French. Humans must retain final control over decisions supported or affected by AI. This is particularly important when AI systems can affect access to rights, public services, employment opportunities, education, healthcare, credit, information, or democratic participation. Opacity cannot be accepted as an inevitable feature of innovation. Brazil supports governance frameworks that promote algorithmic transparency, accountability mechanisms, independent auditing, due diligence by governments and private companies, and effective safeguards against discrimination, manipulation, and widening of social and economic inequalities. Interoperability must not be confused with regulatory uniformity. The compatibility of governance free approaches should enable dialogue among different national policy frameworks. It should not restrict the right of states to adopt their own regulatory regimes according to their national priorities. Digital sovereignty is not digital isolation. It is the ability of states to participate in the global digital economy while preserving their regulatory autonomy and ensuring that data infrastructure and AI systems serve the public interest. International corporations should promote technical interoperability, normative convergence where appropriate, capacity building, and mutual trust while recognizing different national and regional realities. Levels of technological development, and institutional capacities. And the United Nations has a central role in this effort. It can help avoid fragmentation, identify convergence, connect existing initiatives, and ensure that developing countries participate not merely as rule-takers but as full co-participants in shaping the rules of AI To conclude, safe, secure, and trustworthy AI will only be possible if trust is built not only into systems but also into international regimes that govern them. Thank you. Co-Chair [1:26:06]: Obrigado, Brasil. Now we call on Sage Bionetworks, followed by Friends and Witness. Okay, we now call on the distinguished delegate from France, followed by Witness and Botswana. France [1:26:43]: Thank you very much. Distinguished colleagues, co-chairs, the acceptability of AI will depend on the trust that we place in it. There is a major issue to ensure the technology, it does not develop outside of any regulatory framework. Human oversight is the responsibility, and responsibility in AI models are political issues because technology itself does not stop. We are at the point where AI models are being withdrawn from the market because they are deemed too dangerous to put in anyone's hands. We must be diligent to ensure that regulatory frameworks nationally and internationally are not distanced from economic development. France has taken stock of this issue, tackling it during the FII and G7 summit. It's time that we continue our work on this, on this topic. Regionally, the European Union has started to tackle this issue, particularly by adopting its AI regulation. Private actors, for their part, need transparent, a transparent framework and information for their use of AI. For that to work, we must avoid fragmentation and promote the adoption of common practices. Here, under the auspices of the work of France within the G7, 56 enterprises signed up to the principles of the Hiroshima Process on AI piloted by the OECD, which targets common principles and practices for the governance of advanced AI systems. On national security issues and economic— national economic issues, we also have the safety and security of use for our citizens, particularly the youngest among us. AI has strong potential for access to education, but the protection of our youth against risks in mental health, cognitive development, and learning is a priority of France and many other states. It must therefore be a priority for this dialogue among the international community. The issue is also high stakes with regard to the integrity of electoral processes threatened by disinformation disinformation practices that often make use of AI. The multi-stakeholder approach for, for this Global Dialogue on Governance is essential, as is the integration of science and research. Governments must be able to lean on work undertaken in existing multilateral spheres— in the OECD, ITU, the Global Partnership for AI, and the AI Summit. UN organizations must play a frontline role in drawing up safety and security standards that are common to us, then disseminating them among states. We must also collectively tackle the issue of independent evaluation of AI models before their rollout. Thank you. Co-Chair [1:29:49]: Merci, France. We now have Witness, then the distinguished representative from Botswana, and then the Center for Responsible AI at the India Institute of Technology. WITNESS · Executive Director · Sam Gregory [1:30:07]: Thank you, co-chairs. Uh, my name is Sam Gregory, and I'm the executive director of WITNESS. The independent scientific panel is clear: AI is eroding our shared sense of reality. At WITNESS, where we work with both frontline civil society and media, as well as on policy and standards to defend reality in the age of AI and contested truth, We see this directly through our Global Deepfakes Rapid Response Force and in work with human rights defenders and journalists worldwide. We see synthetic clones impersonating public figures, faked scenes from conflict zones, and relentless non-consensual sexual imagery targeting women and girls. And we see the liar's dividend, the panel names real events waved away with the mere claim of AI. So let me address this session directly. Trustworthy AI is impossible without a trustworthy information environment, and that erosion is a systemic safety risk, not just an ethics issue. When we say safe, secure, and trustworthy, the instinct is to reach for catastrophic risk, but the most widely distributed harm is already here: a world of ambient doubt and constant attack, corrosive to rights, justice, and to democracy. When people can no longer tell what is real, every other safeguard we build rests on sand. That deserves attention as the systemic cross-border risk it is. So, two outcomes I would urge for the co-chair's summary. First, read this track broadly. Information integrity must be a focus for interoperability. That includes a focus on meaningful and interoperable authenticity infrastructure and content provenance so we can all read the recipe of the content and communication we consume. It's mix of AI and human. Standards like the C2PA specification must be built with privacy and human rights at their center so they cannot become surveillance infrastructure. Alongside that, we need detection capabilities for deceptive AI that work globally. Our Witness Tried benchmark shows that detection currently works least well for those most at risk, particularly in the global majority. Both must rest on independent third-party oversight. Trust verified by third parties, not asserted by the developers who profit from them. Second, anchor interoperability of the frameworks we already have. We need convergence around existing human rights conventions, including the Council of Europe, international frameworks including the Hiroshima Process, and national and regional legislation, including the EU AI Act. Momentum on these measures to protect our shared understanding of reality in the age of AI already exists across jurisdictions and UN bodies, including the ITU. The task is to make it genuinely interoperable. And a concrete action for governments after Geneva, a concrete locus for interoperability: mandate a multi-stakeholder working group on content authenticity and information integrity on the road to New York. And include clear civil society contributions to that. We have the chance for collective action to defend reality. Please take it. Co-Chair [1:33:21]: Thank you, witness. Now we call the distinguished delegate from Botswana, followed by the Center for Responsible AI at the India Institute of Technology, and Togo. Okay, we're now call the representative for the Center of Responsible AI with the India Institute of Technology, followed by Togo and Mastercard. CRAI [1:33:58]: So, good morning everyone. And I am here representing my colleague Geetha Raju from the Center for Responsible AI who couldn't make it. So, let me make it clear again, I'm not talking on behalf of the panel. And you can just look at me and think I'm Geetha. So, one of the main contributions that we would like to make to this discussion is the urgent need for structured AI incidents reporting. As a core pillar of global AI governance. As AI systems are increasingly deployed in high-stake public sector contexts, the gap between innovation, implementation, and accountability continues to widen. Our work at the Center for Responsible AI focuses on developing AI incident reporting framework that enables the systematic capture, classification, analysis, response, and where possible, resolution of real-world harms alongside clearly defined institutional responsibilities. So, we believe that any incidents framework and any incidents reporting framework has to be democratized at the grassroots level. While AI developers are called upon to provide AI disclosures, right, to support public awareness of AI use, our framework further ensures that no individual or community is left behind, recognizing that AI systems are developed and deployed across multiple stakeholders,. We also recommend establishing shared responsibility and accountability throughout the AI value chain. As an evidence-based policy research organization, CRI strongly emphasizes that such a framework generates real-world evidence of harm, enabling the AI industry and governments to develop feedback loops for better system design, appropriate safeguards, and informed policy responses, and enable us to move from reactive governance to proactive learning-oriented approaches that can keep pace with rapidly evolving AI technologies. So just a side note apart from this, the very fact that I have to present and not Geetha is again demonstrates the difficulty of people from the third world having access to forums such as these. She could not get a visa at such short notice that was given to her to come in. So it is a challenge for even people from the third world to participate in these discussions. Discussions. Thank you. Co-Chair [1:36:20]: Thank you very much, and thank you for that important last message. Before I turn to the next speaker, I'd just like to acknowledge and recognize that we have been joined by the two co-chairs of the International Scientific Panel, Yoshua Bengio and Maria Ressa. Thank you so much for being with us, and thank you so much for the hard work of the panel leading up to the dialogue today. Now I would like to call on the distinguished representative from Togo. After Togo, we will have MasterCard and South Africa. Okay, we will now move to MasterCard. Caroline? No? Okay. The distinguished representative from South Africa. Okay. South Africa [1:37:40]: Tjepersin, Excellencies, distinguished delegates. It's an honor for South Africa to participate in this important dialogue which discusses issues that are central to economic growth, social inclusion, and democratic resilience. South Africa approaches AI governance from a developmental, human-centric, and inclusive lens. Our objective is clear: to harness AI as a catalyst for economic transformation while ensuring it is ethical, safe, and aligned with our constitutional values. There's no doubt that many governments have moved quickly to adopt AI governance frameworks. This is welcome, a necessary step in of view. However, we must be honest. AI has both improved safety and increased complexity. We are seeing greater awareness of risk, including bias, misinformation, and cybersecurity threats. There are also emerging norms around transparency, accountability, and human oversight. However, the pace of AI advancement has far outstripped policy cycles. Governance is fragmented across jurisdictions, often creating overlapping or conflicting rules. The legitimate rise of digital sovereignty pursuits risk creating regulatory silos and uneven standards. For South Africa, the key challenge is not simply governance, but better coordinated, adaptive, and inclusive governance. Many AI governance frameworks such as Hiroshima AI Principles, OECD AI Principles, EU AI Act, etc., have played a critical role in establishing shared baseline, particularly around safety, risk management, and responsible development. However, the AI landscape has evolved significantly. AI is being integrated into sensitive sectors, including defense and national security. The scale and speed of deployment have increased dramatically. This is why you must ensure that global framework reflect not only the priorities of advanced economies but also the realities and needs to developing countries. And to achieve this, it requires that global governance frameworks are inclusive in their design, equitable in their application, and supportive of technology transfer and capacity building. Otherwise, governments risk reinforcing existing inequalities. In closing, South Africa remains committed to working with global partners to advance safe, ethical AI, promote inclusive digital development, and ensure that AI serves humanity as a whole. Thank you. Co-Chair [1:40:35]: Thank you, South Africa. Now we call the distinguished delegate of Oman, seguido de la Mexican Society for Artificial Intelligence de la Universidad Panamericana. Seguimos con— Mexican Society for Artificial Intelligence [1:40:58]: we then continue with Mexican Society for Artificial Intelligence from the Pan American University. Your Excellencies, distinguished guests and colleagues, ladies and gentlemen. There is no use of AI systems without risk. Every deployment involves trade-offs between risks, impacts, and benefits, and these risks, impacts, and benefits are not distributed equitably. As evidenced over recent days, the primary benefits are concentrated among just a few, whether children and vulnerable populations bear disproportionate risk and severe consequences. In response to this, countries adopt different approaches to this false dilemma between innovation and regulation. Similarly, local governments, businesses of all sizes, developers, and consumers find themselves facing the same perceived conflict. I believe that this Global Dialogue is doing crucial work by discussing common tools of international legislation that can help us shape the future for our children and grandchildren. The safety and security of this generation and future generations, as well as trust in AI, can only be achieved if governments reach agreements to protect human rights and dignity of the individuals, and above all, put in place the infrastructure to enforce the law. For any company or, or individual interest in developing or deploying AI systems, navigating this labyrinth of international, local laws, regulations, standards, and principles becomes complicated. In this context, a new international instrument on the governance of AI must prioritize interoperability and a set of legally binded rules that establishes the minimum mandatory standards that are truly applicable in all countries across all regions. Additionally, as an academic from Mexico, a developing country, I believe that equally important is the responsibility that the work of this panel entrusts to us all present. The fight to fight each and every one of us from our own positions to raise awareness, educate, and advocate for AI to be designed, developed, deployed, and used taking ethics into account for the good of humanity. I hope we all leave this event with the motivating— the motivation to continue working to help make AI safer and more reliable. Thank you very much. Co-Chair [1:44:18]: Thank you. We are now going to transition to the next panel, which I will turn over to my colleagues. But just to let you know who will be up on the list of interveners following that panel, we have Japan, the AI safety from UAE, Iran, The International Electrotechnical Commission, Pakistan, Concordia AI, Bangladesh, the 2026 International AI Safety Report, Australia, the United Nations Regional Coordinator Offices, Bulgaria, Vietnam, Lithuania, Global Cities Hub, and Ethiopia to conclude. If you are here, please let the Secretariat over here know that you are in the room so we can make sure we have time for you on the speakers list. Thank you. And thank you to all of our interveners who just spoke. We really appreciate and value your contributions. Facilitator [1:45:16]: Thank you so much. Um, we would also like to add our thank you to the rich interventions. Actually, before we go to our second panel, we would like to shake it up a little bit and get moving around a little bit, because really in the age of AI, we thought we would go back to basics. So let's get away from the machines and get back to the humans. Each and every one of you, turn left, turn right, turn to the front of you or to the back of you. Most importantly, find a person you do not know. So no cheating and talking to someone who who you're already familiar with. Take a moment to introduce yourself, and then you have 4 minutes— we will time those— to discuss one very important question. And please take note of the question before you go and introduce yourselves. If you could— the question is this: if you could identify only one key point or one key priority on safe, secure and trustworthy AI interoperability and compatibility approaches. What is that one key point, or what that one key priority that you would like our co-chairs to report back on? Your 4 minutes start now, and no cheating about talking to someone you already know. Off you go. Oh, thank you. Okay, your 4 minutes are up. Thank you so much, everyone. I hope you had a productive chat with your new friends. Now it's time to turn to a second moderated panel, and in order to guide us through this panel, I'm very pleased to welcome our moderator, Ms. Rachel Adams, the CEO of the Global Center on AI Governance. She will introduce the rest of the distinguished panelists and she will then lead the discussion. Ms. Adams, the floor is yours. Moderator [1:52:03]: Good morning, everybody. It is such an honor to be here today, and it's a pleasure to be moderating this panel on interoperability, building blocks for the global governance infrastructure. Our first panel today helped frame the existing landscape around principles and standards and best practice, and in this discussion we want to move from the what to the how. So the question we will be exploring today is what must actually be shared across borders: evidence measurement systems, evaluation methodologies, reporting tools, institutions' practices so that safe, secure, and trustworthy AI can be governed in ways that are compatible and inclusive. Equally important, we want to avoid a model of interoperability that simply exports assumptions and risks and priorities from powerful jurisdictions. Interoperability must allow for different national contexts and priorities. Our panel will focus on three, three things. First, what common foundations are already emerging? Second, what practical steps governments, industry, and standard bodies can take now? And third, how we can ensure countries and communities with fewer resources are active shapers of global governance infrastructure. So I'm delighted to introduce our panel. Minister Seema Mulhotra, Minister within the Foreign Office of the United Kingdom. Celeste Saulo, Secretary General of the World Meteorological Organization. Dr. Joy Buolamwini, founder of the Algorithmic Justice League. Ravan Samboopille, co-founder and CEO of Credel AI. And Xinghua Lu, who is a member of the AI panel. So thank you all for being with us, and I invite you to join me up here. Minister Malhotra, I'd like to begin with you. The UK has played a very important role in developing AI evaluation methodologies through the UKAC, as well as through international partnerships on AI safety standards and toolkits. From the UK's perspective, What shared evidence measurement standards and evaluation methodologies are most important for establishing a common foundation for safe and secure and trustworthy AI across borders? And how might this work be advanced in ways that respect different national contexts? United Kingdom of Great Britain and Northern Ireland · Minister · Seema Mulhotra [1:55:37]: Thank you very much, Chair. And it is great to be here at the first UN Global Dialogue work on AI and also to be sharing in this important conversation on safe, secure, and trustworthy AI. I, I thought it may be helpful in the context for answering this question to share some of the work that the UK is doing on this and how we see some of those principles being applied. It is extremely important to recognize that AI can be and must be a powerful force for growth, for sustainable development, and for inclusion. But that will only happen if we see countries working together to build trust, to manage risks, and also to ensure that the benefits of AI are shared widely. In that context, it is extremely important that we are thinking about interoperability interoperability and standards. And I wanted to share some of the work that we're doing through the AI Security Institute. Um, the AI Security Institute helps build scientific foundations for trustworthy AI, and it does that through its work on evaluation methodologies, testing, safeguards, and risk assessment. And it is through that helping create a common evidence base for understanding the capabilities and the risks of advanced AI systems. Uh, importantly, this is not the work of the United Kingdom alone, because through the growing network of, uh, AI safety and security institutes, countries are increasingly collaborating on evaluation approaches, on technical research, and shared understanding of frontier AI risks. We've also helped establish the International Network for Advanced AI Measurement, Evaluation, and Science, which we currently chair, recognizing that trustworthy AI must depend on robust measurement and testing and evaluation capabilities. And those capabilities are important, not just the simple high-level principles. And within that, um, that the initiatives that we are working on are seeking to drive evaluation methodologies and reflect that shared challenge that that's facing all countries. We are also the secretariat to the International AI Safety Report, which is also a key piece of scientific analysis alongside the UN Scientific Panel's report. And it is very important for those foundational building blocks that we understand what the state of science is and how that supports policymakers. And I'll just finish by a comment about locally led AI, because one of the lessons we must learn is that it's important that we work in partnership with countries. It's important that we are working inclusively on those partnerships too, because AI must be tailored to local contexts, and that is key to building that trust in AI that is going to be the very important foundation for how a very, very fast-changing space in technology is moving forward and must do so inclusively and with that trust as a foundation. That's why we're very proud of the work we're doing with partners through the AI for Development program. We're looking to build AI tools in more than 40 African languages. We're supporting 13 AI labs, and we're helping countries develop local locally led AI ecosystems that reflect their own priorities. This, alongside the work we're doing to support data security and standards, reflects a wider principle that countries should be empowered to harness AI in ways that support their own development goals. And I'm grateful for all the conversations we're having, we're having here. It's extremely important we see this as a very significant moment with the three dialogues and conferences here in Geneva, that we are looking not just to enhancing our understanding of global AI governance and principles, but how that's applied equally and fairly and inclusively, and including the Global South very centrally in those conversations. Moderator [2:00:25]: Thank you very, very much, Minister. Secretary-General Saulo, I'd like to turn now to The World Meteorological Organization has deep experience in building international cooperation around shared evidence and measurement systems and, and scientific standards. From that perspective, what lessons here might be most relevant for AI governance, and what does it take to build confidence in shared evidence and evaluation systems across borders, recognizing different capacities and contexts. WMO · Secretary-General · Celeste Saulo [2:01:01]: Good day to everyone. Thank you very much, moderator. And certainly, I think that we can use some keywords here, and following what has been expressed by Her Excellency, interoperability, standards, and trust. And that's possibly why you invited WMO, because we've been exchanging data data across borders for more than 150 years. So to exchange data that is usable and— and by different communities across the world, 193 countries actually, you need to have standards and you need to have trust also on that. So this free data exchange builds a lot of experience and now at the— at the— at the table of AI. And that— those are the lessons maybe I would like to share today with you, because trust does not come from just speaking about trust. You need to build trust. It's not something that happens all of a sudden. And, um, going into what AI brings, of course AI is bringing, uh, impressive tools to the area of of weather forecasting and weather observing systems. Now models that needed hours to run can be run in only minutes, and that is affordability. That brings equity as well. But at the same time, of course, it brings also benefits because we are better positioned to forecast tropical cyclones, sand and dust storms, severe weather. But this is part of the exercise. We've been speaking about human-centered approach. And because you build trust only with human beings. So it's how do you build that trust out of a forecast? It's just if you check that the forecast work well. And that the key word for that is Verification. WMO has also a long story of verification because unlike many other systems, we can check with the reality what has actually happened. So we have a forecast, we have a reality, and then we can set also standards to define the quality of the AI-driven system or the dynamical-driven system. So that's the second point I want to make, interoperability. Of course, standards. Of course, verification. But in order to have good verification mechanisms, you need data. You need data on the ground to make sure that your forecast can be shared and compared with the real truth. And then it comes the other question in terms of the new forecast. Are they reliable? Are they good enough for extreme events, for example, or just Are they good enough for sand and dust storms? It's not the same for everything. And that's why we emphasize the need to adjust it to every single location, region, and country. And that is something also mentioned by, by, by the previous speaker. Let's say that we need to construct metrics across the communities to make sure that the AI-driven new algorithms are working properly. And let me just finish by highlighting the results of a recently survey we did across our members. Our members are national meteorological and hydrological services, to, to ask them what is, uh, the main— your main opportunity and concern about AI. What they say is equitable access. That is their main concern, equitable access. So back to the beginning of what you asked to me, we do not need to reinvent the wheel. We have institutions like the WMO that has a lot of expertise in sharing data standards in a trustworthy manner, and that can build from every single country the capacity they need to benefit from AI. Thank you very much. Moderator [2:05:19]: Thank you so much. Thank you. Dr. Liu, I'd like to turn to you now and explore the perspective of technical standards and evaluation and where you see the kind of greatest opportunity for international alignment. What, what elements of AI governance seem most amenable to common approaches like benchmarks, evaluation protocols, incident reporting, and so on, and which elements might need more contextual adaptation? UN AI panel · Member · Qinghua Luo [2:05:52]: Thanks. Hello, everyone. My name's Qinghua Luo. I'm a UN AI panel member and also a researcher from CSIRO Australia. So first, I think the countries need to agree on a common set of risk management principles. For example, those principles can be found in the existing or the proposed legislation. The AI system should be evaluated through the system lifecycle, including during the training stage or after the training, before deployment, or post-deployment through the monitoring stage. And another example principle could be independent third party need to be part of the evaluation process to avoid bias. And if the risk is above the threshold set by the government, the system shouldn't be allowed to be deployed. And second, I think we need to have common methods for evaluating AI systems and also share the results across countries. So, as Mandy mentioned this morning, so there's a shift from static evaluation to— or static governance to dynamic governance. So, benchmarking itself can provide an initial understanding about the AI system's performance. Under like representative configuration, but it's not sufficient, especially now comparing to traditional software, the agent systems. They receive human requirements, human instructions at runtime so they can like have different settings for accessing the tools or accessing the data or different guardrail design. So we need to also have the post-deployment evidence from real-world use. So, this system-level evidence, they really, they can provide more informative information than the benchmark scores itself. And then I think we also need scientific measures to determine what kind of evidence we need to collect and how this evidence should be collected and how commercial confidentiality or private privacy can be protected. And how the evidence, they can be aggregated, analyzed across borders, across organizations. And countries should also be able to collect locally relevant deployment information and using some common evidence model, the trusted organization like AC, they can aggregate scientific insights internationally while preserving the confidentiality. And also respecting countries' circumstances. And in terms of international collaboration, I think we should consider two ways. One is horizontal, one is vertical. So the horizontal approach could be focused on mapping or connecting existing tools or standards. And the government, industrial research organizations, they could share their current evaluation tools or measures evidence through the common platforms. Then the dedicated working groups, they could be focused on building the connectors, like developing mapping between different risk management frameworks or incident taxonomies, evaluation metrics. Then these mappings could be organized based on lifecycle stages or system layers, stakeholder groups, or the principles. Then the Vertical approach could be focused on the collaboration around specific technical challenges. One good example is the International Network of AI Safety Institutes and its joint testing exercise on large language models, multilingual capability testing. So we use a common testing methodologies. Well, every single country, they focus on different languages. Then the results were then shared and compared across countries. So, in this way, we have a richer evidence base than any single country can produce alone. So, that's the two ways, I think, in terms of collaboration. So, it's— in my opinion, it's important to build a shared scientific and technical evidence base to support the country's governance interoperability, and we don't need to, like, build duplicate infrastructure for this evaluation. Yeah. Moderator [2:10:14]: Thank you very much, Dr. Lu. Ravn, I'd like to bring you in here and turn to the practical implementation challenges. Where is fragmentation or the lack of a common evidence base already creating challenges? For organizations that are trying to implement AI governance? Perhaps you could share some real-world examples from your own work to illustrate for us. Credel AI · Co-Founder and CEO · Ravan Samboopille [2:10:38]: Thank you. Yeah, absolutely. Thank you very much for the question. Can you, can you go? Okay, here we go. Um, mic on. Yeah, thank you very much for the question. Uh, I think, as you said, I sit in the implementation layer, which is sort of where, you know, the theory of AI governance gets translated from, you know, documents and principles into like actual code, right? And therefore reality. And we absolutely see both the challenges of fragmentation. But I'd also like to say that we also have like really good cause to be optimistic about the ability for us to solve those fragmentation challenges. We've seen concretely both the challenges and the benefits here. In the Model Context Protocol technical specification. So that's a spec that for those of you who aren't familiar, roughly 2 years ago, which is when Credo was getting started, if you were building an AI system that you wanted to integrate with, you know, other tools that you had at your organization, every system that you wanted to integrate your AI with needed a custom integration, a custom sort of software buildout. In order for that AI system to speak to your other tool. And about 18 months ago, the Model Context Protocol was published. And in those 18 months, it's gone from sort of an idea, an announcement to what is absolutely the global de facto standard that is adopted by every large AI research lab, every enterprise, every AI company like ours as well. And the speed of adoption of that spec has really come from a few simple principles about how the spec was, was designed. And we'll see how it both created and solved some of those challenges. So particularly the— probably the most important learning is that that spec was really designed around the pain points that people at the frontier doing the implementation of these systems were experiencing. Those pain points exist both at the governance layer but also obviously at the functionality layer. But if you design your technical specifications and your standards to solve those problems, then you can do, like, really, really well. Some of the examples in the MCP spec specifically, we've made a lot of progress solving authorization challenges. How do you tell if an AI system is authorized to take an action or not? How do you elicit human approval or human oversight? Into the sort of toolchain of an AI system. These are things that were completely unstandardized, had no norms, no standards around them just 18 months ago. And today, almost every AI system in the world is either built on this standard or beginning to conform to this standard. Now, of course, I would love it if we developed the right authorization frameworks into the MCP spec from day one, which, you know, candidly, we did not. And as a result, we paid the price of having to retrofit the MCP spec to the reality that we discovered after the fact. But I will say that, you know, it highlights both the challenges that, you know, that occur when you get it wrong, but also the opportunity and the clear demand there is in the industry today to develop these governance standards and to develop these authorization frameworks that allow us to create conformity, because that not only accelerates the benefits of AI in the industry, but it actually also increases the governance standards of the average AI system because it gives, you know, everyone in the world the ability to copy these reference examples that are open source published with, you know, running code on day one. And that really accelerates the sort of industry as a whole. So I suppose to like wrap My kind of conclusion there is very much that these open-source standards will very much be adopted very fast by the industry as long as they meet those three criteria of one being oriented around the real-world pain points of deploying these systems at the frontier, especially on the governance side where we sit. You know, regulated institutions like hospitals and government agencies simply cannot deploy these systems unless they have governance frameworks they can trust. Two, they have to be open source so every country and every player can contribute and learn from those reference examples. And three, they have to actually accelerate both the benefits of AI and also the governance frameworks as well. Moderator [2:15:09]: Thanks. Amazing. Thank you so much for sharing all of that. Thank you. Dr. Bulamwini, your work has shown so powerfully that what gets measured and what doesn't get measured has real consequences for people's rights and opportunities. So, as countries and institutions develop shared evaluation systems and standards for trustworthy AI, what must we ensure gets included so that these systems capture bias and discrimination rather than simply making AI easier to deploy? Algorithmic Justice League · Founder · Joy [2:15:44]: No, thank you for having me and the opportunity to speak here today. I am Dr. Joy, founder of the Algorithm The Algorithmic Justice League. I will actually start with a poem to ground us back to why we're here, which is for the X-coded, the people who are exploited, condemned, or otherwise harmed by AI systems. So it is called "Unstable Desire." Prompted to competition, where be the guardrails now? Threat in sight will might make right. Hallucinations taken as prophecy. Destabilized on a middling journey. To outpace, to open chase, to claim supremacy, to reign indefinitely. Haste and pace control altering deletion. Unstable desire remains undefeated. The fate of AI still uncompleted. Responding with fear, responsible AI beware. Prophets do stare and people still dare to believe our humanity is more than neural nets and transformations of collected muses. More than data and errata, more than transactional diffusions, are we not transcendent beings bound in transient form? Can this power be guided with care, augmenting delight alongside economic destitution? Temporary Band-Aids cannot hold the wind when the task ahead is to transform the atmosphere of innovation. And so that's what we're here to do when it comes to transforming the atmosphere sphere of innovation. And when it comes to what evidence is necessary and how that evidence is collected, one core part that I continue to see missing when it comes to AI harms reporting and incident reporting is actually to include the people who themselves are being harmed. So, when you submit an incident harm report, it's not just the model, though, that it's important. It's the person and understanding those demographics so later on when we move from just looking at how we document harm and looking at how we investigate harm but looking to remedy harm, we've already architected that throughout the entire system. The other area to avoid would be misleading measures of success. With my foundational research on algorithmic auditing, many of the gold standards from institutions like the National Institute for Standards and Technology, which have senses improved, oftentimes did not represent the global south, did not represent people with dark skin. And as a result, we thought we were making progress when we then inspected the data by which the standards themselves were set that were gold standards. They proved to be pyrite. Oftentimes less than 5% of a benchmark might represent the global majority. And that's not a mistake we can afford to make moving forward into the future. So let's learn, let's do it better so we can serve the X-coded and transform the atmosphere of innovation. Thank you. Moderator [2:18:44]: Thank you so much for giving us such a profound reminder of our purpose here. And thank you to all our panelists for their comments and inputs, and all the best with the deliberations today. Thank you. Facilitator [2:19:40]: Thank you once again to our moderator and the panelists of Panel 2. We now hand the floor to our co-chairs to manage the the floor once again for a second round of interventions from member states and other stakeholders. IAHWG · Co-Chair [2:19:59]: Okay, now we now call, um, Japan, followed by AI Safety UAE and Iran. The distinguished delegate for Japan. Thank you. Japan [2:20:50]: AI is the cornerstone of Japan's growth strategy and is expected to serve as a driving force for global economic and social development. In order to promote innovation through AI and to drive economic and societal growth, it is essential to build a safe, secure, and trustworthy AI ecosystem. It is important that we share an interoperable ecosystem that minimizes the risks posed by AI, including those related to information security. The key to that is trustworthiness. Based on this building principle, Japan took the initiative to launch the Hiroshima AI process and has been taking the lead in international discussions on AI. Through the Hiroshima AI process, efforts toward inclusive international governance on AI have been advanced through such concrete initiatives as the development of guiding principles and a code of conduct as the first international norms for generative AI. Today, approximately 70 countries and regions, including the Global South, have endorsed these principles as the Friends Group and share knowledge and experiences related to AI government and policy in order to achieve trustworthy AI. Leading global companies at the forefront of AI innovation are also partners in this process. For example, the reporting framework promotes transparency and accountability in how private companies develop generative AI. In this way, the Hiroshima AI process has been contributing to the advancement of international AI governance in cooperation with the United Nations. We are confident that it also offers valuable perspectives in the context of this global dialogue as one of the existing best practices. Japan will continue to lead these efforts and contribute to co-creating a safe, secure, and trustworthy Ecosystem. Co-Chair [2:23:11]: Thank you. Do we have the AI Safety UAE in the room? Nope. The distinguished representative from Iran? Okay. I would like now to call the International Electrotechnical Commission. They will be followed by the distinguished representative from Pakistan and then Concordia AI. IEC [2:23:46]: Yeah. Thank you, co-chairs. I represent the IEC. The International Electrotechnical Commission. We are very grateful for the recognition given over these two days to the importance of international standards, including those we develop with ISO. The recognition is also reflected in the preliminary report of the scientific panel. At the IEC, we find it useful to think of AI governance along five normative layers: international law, international soft law, national law, international standards, and the normative power of what is technically feasible, basically shaped by the industry and the private sector. International standards play a distinct role in this landscape. They form the link between these layers. They translate the principle set out in international law and soft law into practical tools. They give governments a basis for national regulation, and they give industry a common language to build safe and interoperable products. The value of our systems lies in who takes part in it. International organizations, industry, national governments, the private sector, academia, civil society. They all have a voice around the table, and they have a voice in developing international AI standards. This happens through our members, one per country. It is a system built on consensus and open to all. We consider international standards as an important piece of the puzzle for global AI governance. However, we recognize that we are not the only piece. There are many other pieces. The picture is larger, and it requires us to work all together. So it is with this understanding that we come to this dialogue. We wish to contribute meaningfully to a form of AI governance that delivers what we may call AI with trust. We are ready to work with all of you so that AI systems remain safe and trustworthy across borders. Let me add one last point. A core expertise of the IEC lies in electricity. This gives us a particular contribution to make AI much more energy efficient, in addition to be trustworthy. Through standards for energy efficiency, data centers, grid resilience, and to support access to electricity and energy security, we can make infrastructure that is resilient and bankable. Again, without reliable power, there is no AI. Thank you very much. Pakistan [2:26:50]: Thank you. We now call the distinguished delegate of Pakistan, followed by Concordia AI and Bank Bangladesh. Thank you, co-chairs and co-leads. Excellencies, ladies and gentlemen, colleagues, a very good morning. Pakistan welcomes you to this cluster and welcomes this cluster. Safe, secure, and trustworthy AI is a shared interest of all nations, but we must be clear about how safety is defined, by whom, and for whose interest. Let me offer 3 points. First, Safety must be defined inclusively. Too often, the standards and benchmarks for trustworthy AI are being set by a small number of states and firms through national safety institutes and frameworks in which most of the world has no voice. Safety is not purely technical matter. It is a contextual, shaped by language, culture, and local realities. If the developing world is absent from the table where these norms are made, the result will be standards that are exported rather than shared. Pakistan therefore calls for safety norms to be developed through inclusive multilateral processes under the United Nations, where all nations participate as equals. Second, interoperability must connect. We support the interoperability and compatibility of governance approaches as agreed in the Global Digital Compact. Interoperability allows trade, research, and data to flow across borders and lets diverse national approaches coexist. But it must be a bridge between systems, not a mechanism through which the rules of a few become binding on the many. Common standards should be built through open collaborative standards bodies upholding safety, reliability, and human rights, and reflecting linguistic and cultural diversity throughout the life cycle of these systems. Third, trust requires capacity to verify. A nation cannot trust what it cannot assess. Yet the ability to test, evaluate, and assure AI systems remains concentrated where the technology is built. Without deliberate effort, Developing countries will be asked to accept assurances of safety they have no means to verify for themselves. We must therefore treat the capacity to evaluate and govern AI as an integral part to safety itself, supported through shared tools, pooled expertise, and cooperation across regions. Underlying all these three is a single principle: AI systems must be developed and used in full compliance with international law and the United Nations Charter, with meaningful human oversight in both civilian and security domain. Pakistan's National AI Policy tries to be comprehensive and encompass all these aspects, including that of ethical AI deployment. AI sovereignty, especially in terms of the infrastructure that, that is Capital intensive is something which is going to be integral for all the trust that we want to share. It is going to be extremely important that we collectively invest and make sure that no city, no country, and no citizen is left behind. Thank you. Co-Chair [2:30:50]: Thank you. We will now call Concordia AI, followed by the Distinguished Representative from Bangladesh, and then the 2026 International AI Safety Report. Concordia AI [2:31:05]: Thank you, co-chairs, distinguished delegates, and colleagues. For the past few years, our governance challenge has been about what AI chatbots said. In 2026, that challenge has fundamentally shifted. We must now govern what AI does. In real time. If there is one defining narrative for this year, it is the rise of agentic AI, driven by explosion of frameworks such as Open Claw that are transforming AI into active digital assistants. Our understanding of both the opportunities and the risk must be grounded in science. Both the UN Independent Scientific Panel on AI and the International Safety Report reach a critical conclusion: because AI agents can act directly upon the real world, virtually and increasingly in the physical world, their failures can potentially cause severe harm before a human can even intervene. As AI agents become increasingly autonomous in the coming years, the stronger the safety standards must be, similar to how we have standards for autonomous driving from Level 0 to Level 5. And this is especially urgent for agents deployed in safety-critical industries countries, from financial services to energy grids and even in our food supply. And when autonomous systems fail, consequences do not stop at national borders. A malfunction in one country can disrupt shared infrastructure, supply chain, or public safety elsewhere. True international cooperation means ensuring that red lines and safety standards are developed inclusively. Representing the voices of everyone at the table. Now, we must strike the right balance between innovation and safety. For example, through its AI+ initiative, China aims to deploy agentic AI across key sectors in the economy in the next 5 years, while at the same time advancing national safety guidelines and binding standards at a pace of innovation. That is precisely why cooperation between the global scientific community industry is so urgent. In Singapore in May, 100 experts from around the world have found exactly this agenda. The Singapore Consensus identified 10 foundational principles for managing AI risk based on emerging frameworks as well as engineering practices from across different jurisdictions. Now, the implementation of these principles varies in maturity. Some, like minimum privilege, like data lineage or traceable agent identity have workable tools behind them. Others, especially the governance of multi-agents, are still open research questions. One thing is clear: we cannot safely manage a global ecosystem of billions of digital agents without a shared global architecture and standards. The window for proactive governance is narrow. Now, by choosing cooperation, over fragmentation today, we can ensure that the autonomous world of tomorrow remains a world where it is still safely guided by human intent. Let us build that shared foundation together, and thank you all. Bangladesh [2:34:20]: Thank you. We now call the distinguished delegate from Bangladesh, followed by 2026 International AI Safety Report in Australia. Chair, Excellencies, and distinguished delegates. AI is defining the course of the 21st century. Bangladesh believes that safe, secure, and trustworthy AI must must begin with a human-centric approach, one that respects human values, protects rights, and delivers benefits for all. Today's AI systems are increasingly powerful yet opaque. Its decision-making process is uninterpretable. It may generate inaccurate or misleading outputs with confidence and remain vulnerable to adversarial attacks, data poisoning, and cyber threats. Models trained on historical data can also perpetuate bias and discrimination, disproportionately affecting vulnerable groups. Vast amounts of data processed by AI raise serious concerns regarding privacy, security, and confidentiality. These challenges are compounded by fragmented global governance. Divergent regulatory approaches, the absence of common risk assessment standards, interoperability, communication protocols, and globally agreed principles for algorithmic accountability and cross-border data governance create uncertainty. Meanwhile, policymaking continues to lag behind technological advances. As the Secretary-General aptly said yesterday, I quote, "Innovation needs guardrails," unquote. We therefore call for action in the following areas. First, a comprehensive governance framework should combine binding legal obligations with ethical standards supported by an independent global evidence base and continued international dialogue to harmonize governance approaches.. It must uphold human rights, transparency, accountability, privacy, and fairness while effectively addressing algorithmic bias, misinformation, disinformation, and AI-enabled gender-based violence and children abuse, including deepfakes. Second, global AI governance must be inclusive and representative. Standards and regulatory frameworks Solutions should not be developed by a few and applied to many. Developing countries must have an equal voice in— to ensure diverse development realities, languages, and cultures are reflected. Third, humans must remain in the loop for high-impact decisions with clear accountability for AI-driven outcomes. Regulatory approaches should be risk-based and innovation-friendly, including regulatory sandboxes to enable startups and innovators in developing countries to build trustworthy AI without undue regulatory hurdles. Finally, this dialogue should continue to bring people across the table to argue, debate, and discourse to build an AI future that is safe, secure. Thank you. Co-Chair [2:38:01]: Thank you very much. We will now have the International AI Safety Report, followed by the distinguished representative from Australia, and then the United Nations Regional Coordinator Offices. International AI Safety Report · Lead Writer [2:38:24]: Good morning, everyone. I'm the lead writer of the International AI Safety Report. Very honored to be here representing the hundreds of experts who contributed to the report every year. This report exists because governments and organizations and individuals making decisions about AI need reliable, independent evidence to guide those decisions. And to that end, I'm very happy that our report now sits alongside the UN's International— Independent International Scientific Panel on AI, because these two reports are very complementary. While the UN report takes a sort of more holistic view and considers a variety of different kinds of AI, our report goes deep on frontier risks from general-purpose systems in particular, where the evidence is sketchiest and where expert analysis may prove most valuable. Our report this year emphasizes that those frontier risks cannot be dismissed as speculative concerns. You know, we have a range of evidence from laboratory studies to real-world incidents that now show that those concerns have merit. Frontier AI systems are now very capable, and they are still improving very quickly, and grappling with their risks cannot be deferred indefinitely. What are those risks? First, misuse. So, how do we stop people from using AI systems to cause unacceptable amounts of harm? Recently, Mythos brought cyber risks to everyone's attention, but the report emphasizes that other domains need attention too. For example, as AIs get better at scientific reasoning, they are becoming useful expert advisors in various domains, and they may lower the barriers for bad actors to conduct attacks, including potentially very severe ones with bioweapons or other serious weapons. Second, malfunctions. How do we get AI systems to reliably behave as we would like them to? AIs are increasingly agentic. They take many steps towards achieving a goal in the real world with little human supervision. This of course makes them much more useful and more productive, but it also makes them harder to oversee. And because we still understand so little about the inner workings of these systems, ensuring that these more powerful agents stay under human control, that we can correct or stop them if they start behaving badly, is an urgent and persistent problem. And finally, systemic impacts. How will we manage the broad societal and economic transformation that this technology will bring? These risks may prove the most significant, but they are also the least well understood in the report. It is striking how little we still know about how AI systems will affect our work, how we gather information and get advice, and even how we relate to each other and to our governments and how states relate to each other. In fact, across all these risks, our methods for anticipating, measuring, and mitigating harms remain nascent. The existence today of deepfakes, overly sycophantic AI systems, AI-assisted cyberattacks show that we cannot reliably prevent these harms. And as capabilities improve, that risk management shortfall could lead to very serious and, yes, even catastrophic harms. But we cannot manage what we do not measure, and so our report aims to help the world clearly see those risks and what we can do about them. Thank you. Australia [2:41:39]: Thank you. We now call the distinguished representative from Australia, followed by the United Nations Resident Coordinator in Bulgaria. Thank you, co-chairs and distinguished delegates. Australia welcomes the opportunity to discuss efforts to ensure interoperable and compatible approaches so we can accelerate the benefits of AI. AI that is safe, secure, and trustworthy. As we have all heard, these emerging challenges demand nuanced and flexible solutions that support innovation whilst also mitigating the known risks and harms. Our efforts must be multi-stakeholder, industry-informed, and consensus-based. We should build on existing international work rather than creating frameworks from scratch. The International Organization for Standardization, ISO, has developed AI standards that offer a common baseline that can improve consistency, interoperability, and trust across jurisdictions. Initiatives such as those led by the ITU, the OHCHR, and UNESCO have already produced guidance and frameworks for responsible AI. The AI Summit Series, the Hiroshima AI Process, and the Global Partnership on AI have generated practical lessons, policy recommendations, and mechanisms for international cooperation. New efforts must be informed by tested approaches. Regional initiatives such as the APEC AI Initiative, ASEAN's AI Governance Framework and the African Union Continental AI Strategy demonstrate how shared principles can be adapted to different economic, social, and regulatory contexts. A one-size-fits-all approach to AI governance will not be effective. We all face different risks, priorities, levels of digital development, and regulatory environments. Effective governance require— requires shared principles that can be implemented flexibly across different contexts. Responsible AI is not a one-off task, but a continuous process of governance, monitoring, adaptation, and improvement as technologies evolve and risks emerge. Australia believes that the UN Global Dialogue has the potential to bring together the multi-stakeholder community to address AI governance challenges and work together to address them by sharing with each other We strengthen the digital ecosystem, enhance connectivity between our economies, and contribute to a stable, safe, inclusive, and innovative digital ecosystem, importantly, for all. Thank you. Co-Chair [2:44:31]: Thank you very much. I now am delighted to call the United Nations Resident Resident Coordinator Office, followed by Bulgaria and Vietnam. UN · 5 Resident Coordinators serving Angola, Eswatini, Costa Rica, Montenegro, and Bosnia and Herzegovina [2:44:53]: Thank you, Excellencies. Good morning. I speak today on behalf of 5 Resident Coordinators serving Angola, Eswatini, Costa Rica, Montenegro, and Bosnia and Heskovinia. From where we serve, we see both AI's promise and its risks. In Luanda, engineers are building one of Africa's first sovereign AI facilities. Yet countries enter this transition from different starting points. Advanced AI capabilities, computing infrastructure, and investment remain concentrated among a limited number of actors. Inclusive governance must ensure that all countries have the capacity and voice to shape how AI is developed, deployed, and regulated. If interoperability is to deliver development dividends for all countries, some priorities emerge. First, establish a common rights-based floor that narrows the digital divide and protects human dignity. Costa Rica, which ranks among global leaders in algorithmic transparency in government, shows that smaller countries can also help shape global standards. Two, build the capacity to govern AI, not simply to use it. That means stronger institutions with regulators, judges, and civil servants equipped to oversee nationally owned AI systems. Across our countries, AI readiness assessments in Montenegro, national dialogues in Eswatini, and public sector AI initiatives in Bosnia and Herzegovina are already building momentum. The next generation of UN cooperation frameworks can help scale these efforts. Making this a reality requires practical support at country level, bringing together governments, UN expertise, and partners to strengthen institutions, support shared learning, and translate global principles into national action. Excellencies, AI's growing demands for energy and water should also inform discussions at COP31 and at the 2026 UN Water Summit. Interoperability must connect not only systems and institutions, but also people and communities. Ensuring that AI reflects local realities and delivers shared public value. Resident coordinators, together with UN country teams, we convene in 162 countries and territories, stand ready to accompany member states in shaping an inclusive AI future and translating global AI governance into better-lived realities for billions of people worldwide. I thank you. Lithuania [2:47:47]: Thank you. I now call the distinguished delegate of Bulgaria. Not here. Okay, we move on to the distinguished delegate of Lithuania, followed by the Global Cities Hub in Ethiopia. Thank you, Chair. Excellencies, ladies and gentlemen. Artificial intelligence is becoming an extremely impactful technology, but without trust, there will be no meaningful AI adoption. Citizens will not use AI they do not trust. Governments will not deploy it at scale without safety and accountability, and companies will not invest in fragmented and unpredictable environment. For Lithuania, trust rests on three pillars. First, we must ensure that AI stays human-centered and grounded in democratic values. Technology should complement, not replace, human decision-making and responsibility. AI governance must protect human rights, democracy, and the rule of law. Human-in-the-loop principle that requires human oversight, validation, or intervention at critical decision points must be preserved and not treated as an obstacle for innovation. Rather, it is a prerequisite for safe and responsible innovation helping to maximize the benefits of new technologies while minimizing their risks. Translating these principles into practice requires sound governance. In Lithuania, these principles are reflected in our national AI strategy guidelines and are increasingly embedded in our broader digital and AI policy framework. Second, AI must be secured and resilient by design. As AI becomes integrated into critical infrastructure, public administration, and information ecosystems, cybersecurity cannot be viewed as an isolated issue. Lithuania's experience is instructive. As a country that faces persistent cyber threats, disinformation campaigns, and hybrid attacks, We have learned that resilience must be built into technology from the beginning. This is why Lithuania consistently invests in cyber capabilities, critical infrastructure protection, and secure digital government. Our approach is clear: AI deployment must go hand in hand with cybersecurity and risk management and human oversight. Third, and most importantly, we need interoperability and compatibility of approaches. This is why Lithuania strongly supports international AI governance based on common values and practical cooperation. And in this regard, Council of Europe's Framework Convention on Artificial Intelligence, Human Rights, Democracy, and the Rule of Law, the Vilnius Convention, represent a major milestone. Ladies and gentlemen, the future of AI will not be shaped by technology alone. It will be shaped by choices we made about governance, security, and cooperation. Thank you. Co-Chair [2:51:52]: Thank you very much. We next have the Global Cities Hub, followed by Ethiopia. And I believe Dubai Cable is our last intervention. Global Cities Hub [2:52:09]: Distinguished co-chairs, excellencies, ladies and gentlemen, the Global Cities Hub applauds today's discussion and believes that the global dialogue on AI governance represents an important milestone in building inclusive, effective, and future-oriented governance frameworks for for artificial intelligence. The panel's report provides a valuable assessment of the opportunities and challenges presented by AI and rightly emphasizes the complexity of governing this rapidly evolving technology. We heard at the beginning of this cluster this morning that we do not need more principles. We need to translate principles into practice. We are confident that local and regional governments have a role to play in that. As we discuss the future of AI governance, it is essential to recognize that cities and regions are not merely stakeholders. They are public authorities with significant responsibilities across many of the sectors where AI is already transforming our societies. Artificial intelligence is reshaping decision-making in healthcare, education, mobility, energy, environmental management, and public administration. These are policy domains where local and regional governments often hold direct responsibilities and where AI impacts are experienced first and most visibly in cities by two-thirds of the global population. LRG's operational knowledge can significantly improve the evidence base for policymaking by showing how AI affects communities, public services, and urban systems in practice. In this sense, they represent a still missing governance layer which could contribute to the horizontal facilitation of international learning, as we heard this morning as well. Global fragmentation, be it regulatory, accountability, or compliance, exists among cities, as much as among nations. We agree that in public interest, we need to build AI regulation inclusively and operate reliably and securely. Let me conclude with a recommendation. Local and regional governments should be recognized as a distinct governance actor alongside national governments, industry, academia, and civil society. They should be systematically included in thematic discussions on interoperability and compatibility, and in policy discussions and evidence gathering on AI deployment in a safe, secure, and trustworthy manner. For AI governance to be truly inclusive, effective, and people-centered, the voices of cities and regions must be part of the global conversation. Their participation will strengthen the legitimacy, relevance, and practical impact of this global dialogue and help ensure that AI delivers sustainable and equitable benefits for communities everywhere. Thank you very much. Co-Chair [2:55:18]: Thank you. We now call the distinguished delegate of Ethiopia, followed by Dubai Cable. Ethiopia [2:55:48]: Thank you so much, co-chairs. Excellencies, distinguished delegates, as we tried to highlight at yesterday's Bridging AI Divide thematic panel discussion, Ethiopia has enacted relevant legal and policy frameworks, established institutional frameworks on AI, and is carrying out pertinent administrative work accordingly. Mention can be made of our Artificial Intelligence Institute and the planned AI University. It is in recognition of these growing efforts at home that the African Union has designated His Excellency Prime Minister Dr. Abiy Ahmed as Champion for AI and Digital Health, for which we are grateful and honored. So Ethiopia supports the objective of advancing safe, secure, trustworthy, and inclusive AI governance. As AI increasingly transcends borders, interoperability and compatibility of approaches cannot be achieved through isolated national efforts. They require continuous dialogue, mutual trust, and strong multilateral cooperation. Interoperability is not only about creating a single global AI law, about requiring countries to adopt identical governance models, but also about ensuring that different AI government systems can communicate, cooperate, and work together effectively while respecting different legal systems, development priorities, and levels of capacity. To achieve this, we must build a common foundation based on shared principles, common terminology, scientific understanding, and compatible standards. We must also ensure that all countries have equitable opportunities to contribute to and benefit from AI governance. Building the AI Divide is therefore essential if interoperability is to be meaningful and inclusive. For developing countries, fragmented governance frameworks, divergent standards, and Unequal access to technology and technical expertise risk widening existing inequalities and limiting meaningful participation in the global AI ecosystem. So, co-chairs, Ethiopia's commitment to this vision of interoperability and the compatibility of AI systems, governance systems vision, is rooted in our long-term belief in multilateralism and believes in inclusive dialogue, continuous inclusive dialogue under the United Nations. Ethiopia believes that continuous and inclusive dialogue under the United Nations is essential to building trust, shared understanding, and scientific consensus needed for interoperable AI governance. Thank you so much. Co-Chair [2:59:05]: Thank you very much. And our final speaker, if they're here, is Dubai Cable. Dubai Cable [2:59:31]: Thank you, Co-Chair, Excellencies, distinguished delegates, and colleagues. For years, AI governance asked a simple question: how we govern the system that responds to us? Today, The system before us do more than respond, which they act, plan, call other system, and complete multi-step task with minimum human prompting or intervention. That's a shift called now a new approach to the governance. Being part of Enterprise A strategy, building and governing system that moves Goods, data, decision at scale. The strategy without governance is speed without brake. The governance without strategy is a brake with no engine. Once an AI agent is authorized to act, the line between a policy document and the production incident becomes very thin. Two shifts matters most. First one, accountability must be engineered before deployment with the permission, oversight, audit trail need to be designed to build a time-based, not added after an incident. Second one, an agent built under one authority can operate across borders the moment it connects to the global supply chain or financial system. The greatest value of this AI dialogue is enabling our existing framework to work together. In practice, 3 questions apply in every context. First one, what this agent do without, without a human in the loop? Second one, what happens the moment it does something we did not anticipate? The third one, who by name and by role is accountable for the outcome? If a framework answers those questions in plain language, it is ready for agentic system. The agentic AI is not a future risk, it is a present capability already being deployed across enterprises, private sectors, and other areas, ensuring whatever an agent act on behalf of an organization or an enterprise, a clear chain of human accountability stands visible behind it. So that's governance with responsible AI. Thank you. Facilitator [3:02:16]: Wonderful. Thank you so much for all the thoughtful interventions. Before we move to today's summary and closing, I'd like to invite everyone to participate in a second interactive activity. The first activity you all did earlier today, you spoke to each other and you identified a priority item. I hope you still remember that. What I'd like to do is see if you can— there are ITU and other colleagues around who have Post-it notes. Please write down that priority item, and if your group, the few of you who are talking, could not agree on one, each one of you can write one option each, and as you leave, please hand these Post-it notes at the exit of the room, and the co-chairs will take those points into consideration. But with that, maybe I'll hand over to our thematic coach— co-chair— co-chairs to provide a summary of the key messages and emerging priorities from this session. And these reflections will inform the report back to the Dialogue of Dialogues, which will be held later today. Back to the co-chairs, please. PAI · Co-Chair · Rebecca Finlay [3:03:28]: Thank you very much. Thank you to everyone for all of your interventions today. Thank you especially to our panelists for their time and reflections. —and thank you all for taking a moment before you leave today to jot down your priority. We will take that, we will inform the work that we are going to do in summarizing a report back to the co-chairs. And thank you most especially to Minister Bogantes. It has been such a pleasure to co-lead this session with you. Here are a few reflections from what was an an extremely diverse and important set of interventions. And I perhaps want to begin by returning to what Amandeep Singh Gill reminded us in the very beginning, that we are together pioneering a new approach to international learning. Perhaps we are, as Dr. Joy said, transforming the innovation atmosphere in which we work. It's marked by an unprecedented effort today to include and engage voices from across every region of the world to advance, as we are talking about, safe, secure, and trustworthy AI. And my mind is literally still buzzing with the sheer number of ideas which I will not be able to do justice to. Um, let me try quickly to give you 5 key takeaways from the discussion. First of all, safety frameworks are too often developed without The evidence needed to shape them. We heard that from the safety and scientific panel yesterday, but also other efforts to strengthen scientific consensus are so important. Second, we need a shared minimum baseline that builds on existing global frameworks. This is one of the points that my colleague made so clearly. This may not be sufficient on its own, but we need a foundation— to build on, and to a large extent it already exists. We have global commitments, we have international law, we have this dialogue, the Hiroshima Process, the G7, and the body of pledges from the AI summits to date. What we need now is to consolidate, share, and further develop them so that they are implementable and adaptable to regional context. Third, Interoperability is a necessary but not sufficient condition for safety and security. It should rest on essential principles such as human rights. It should be grounded in strong scientific foundations, and it should reflect frameworks developed not only by a few powerful nations but by the world's majority. Fourth, inclusion and openness must be at the heart of our approach. And inclusion is also, as we heard a technical matter. It means open benchmarks. It means ensuring linguistic diversity. It means infrastructure for open science and technical standards that reflect global needs. And finally, trust must be built through verifiability that is fostered through independent third-party testing, disclosure across the AI value chain, including incident reporting, as my co-lead also mentioned earlier. For this dialogue and for international efforts to succeed, two critical aspects I heard this morning need to be reinforced. First of all, multi-stakeholder participation is not a nice to have. It is a necessity for the legitimacy of our efforts and the success of our goals. We act today in the public interest. The dialogue strength rests on bringing together voices from across sectors and regions and just not government and industry alone. That means ensuring meaningful participation. We heard about the importance of ensuring visa processes so that individuals can participate from global communities, and we need equal partners from the start. And this goal needs to be sustained and further supported as we move forward. And finally, we need greater continuity and action between dialogues. Your input that you are providing to us, the interventions we heard today, we will take that and, and inform the co-chair summary report so that what we do today will not pause between summits and restart from scratch each time. What happens between now and the next dialogue and the upcoming AI summits that will also happen here matters as much as what happens the ones in this room today. We will carry forward the responsibility to sustain the momentum and keep this work going. There is— this is a critical moment for all of us to be working together to advance this work. Thank you all so much. Thank you again to Minister Boghantist and her team for their collaboration, and I look forward to your further interventions. Costa Rica · Co-Chair · Paula Bragantes Zamora [3:08:32]: Thank you, Rebecca. I'm going to try to be quick because you summarized it quite well and probably better than me. Today's discussion made clear that effective AI governance will depend not on uniformity but on our ability to connect different approaches through shared standards, evidence, safeguards, and cooperation with preserving national context and ensuring meaningful participation from all countries. I will highlight just a number of key points for building AI governance that is practical, trustworthy, and inclusive. AI governance must evolve from static rules towards adaptive systems capable of responding to increasingly autonomous and multi-step AI. Regulatory fragmentation creates compliance burdens, accountability gaps, and regulatory arbitrage, disproportionately affecting countries with limited institutional capacity. Interoperability should connect different governance approaches through shared technical standards without imposing a single regulatory model or undermining national autonomy. Cross-border regulatory sandboxes can test AI applications under joint supervision before wider deployment. Common definitions and granular mapping between legal requirements, technical controls, and compliance evidence are essential for effective interoperability. Shared incident reporting mechanisms can ensure that lessons from failures in one jurisdiction prevent repeated harms elsewhere. Trustworthy AI requirement requires human oversight, transparency, accountability, and independent evaluation throughout the entire system life cycle. Evaluation frameworks must reflect linguistic, cultural, and demographic diversity, including communities underrepresented in current global benchmarks. Open standards shared Methodologies and executable reference frameworks can accelerate adoption and strengthen international cooperation when they address real implementation challenges. And lastly, information integrity, content provenance, and deepfake safeguards require interoperable approaches grounded in privacy and human rights. Facilitator [3:11:16]: I guess with that, thank you so much to the co-chairs and for all the interventions and the panels. We conclude this session. Thank you so much.