Thematic discussion 3: Safe, secure and trustworthy AI - Global Dialogue on Artificial Intelligence Governance - Day 2

The facilitator argues that just as different electrical systems around the world are made compatible through adapters and common standards rather than uniformity, AI governance faces the same challenge. Countries will continue to develop different laws and regulatory traditions, making full harmonisation neither realistic nor necessary. However, interoperability — enabling different approaches to communicate and work together — is essential.

The facilitator used the analogy of different electrical systems, plug shapes, and voltages across countries, noting that laptops and phones still work safely when travelling not because every country uses the same system, but because adapters, common standards, and trusted interfaces have been developed . The facilitator explicitly stated that harmonisation is neither realistic nor entirely necessary, but interoperability is essential, and that 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 realise .

Paula Bogantes Zamora argues that the world does not need more AI principles but rather a common way to prove they are being implemented. The real interoperability challenge is demonstrating safety, accountability, and trust across borders in practice, not merely having frameworks that coexist theoretically.

She stated directly that 'the world does not need more AI principles. It needs a common way to prove they're being implemented' and framed the challenge as 'not whether frameworks can coexist on paper, but whether safety, accountability, and trust can be demonstrated across borders' .

Bogantes Zamora argues that the extraordinary concentration of AI development in a small number of countries means that those with the greatest computing capacity also have disproportionate influence over global AI governance norms. This includes determining which risks are measured, which benchmarks are accepted, which languages are evaluated, and which institutions are recognised as capable of certifying safety.

She cited that 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 held roughly 75% of the computing power among the 500 largest known AI clusters, with China holding another 15% . She concluded that 'concentration of infrastructure becomes concentration of evidence' .

Bogantes Zamora highlights the severe lack of participation from the majority of the world's countries in international AI governance discussions. This exclusion is compounded by the fact that fewer than 1% of the world's AI clusters are located outside the dominant countries, and one-third of developing countries lack national AI strategies.

She cited a preliminary scientific assessment indicating that 118 countries, mainly from the Global South, are not meaningfully engaged in the principal international discussions of AI governance, while fewer than 1% of the world's AI clusters and one-third of developing countries have national AI strategies .

Bogantes Zamora argues that as AI systems become more agentic — acting on behalf of people or institutions across multiple platforms — interoperability must extend beyond technical system compatibility to include verifiable agent identity, traceable delegation, machine-readable permissions, and effective revocation. Future interoperability must connect not only systems and rules but also identities, mandates, and chains of responsibility.

She cited Estonia's Arawite initiative as an example of this emerging frontier, noting it is developing identity and registry frameworks, interoperability standards, legal recommendations, and a public sector pilot so that states can interact with AI agents in a trustworthy and accountable way . She specified that agentic AI interoperability must include verifiable agent identity, traceable delegation, machine-readable permissions, secure data exchange, auditable actions, and effective revocation .

Bogantes Zamora proposes a concrete and achievable starting point called 'minimal viable interoperability', which focuses on advancing specific practical elements before the 2027 dialogue. This approach prioritises shared terminology, comparable risk classifications, consistent documentation, interoperable incident reporting, and multilingual evaluation methods, while building pathways towards mutual recognition of testing, auditing, and certification.

She explicitly proposed that before the 2027 dialogue, progress should be made on shared terminology, comparable risk classifications, consistent documentation, interoperable incident reporting, and multilingual evaluation methods, while building pathways towards mutual recognition of testing, auditing, and certification .

Bogantes Zamora identifies language as a decisive test for interoperability, noting that the vast majority of the world's languages are excluded from AI training and evaluation. This creates safety risks as models perform less reliably in underrepresented languages, and means that languages from Latin America and other regions cannot remain outside global evaluation frameworks.

She noted that the world has more than 7,000 languages but AI training and evaluation cover only a fraction, and that models often perform less safely in languages with limited digital data . She specifically mentioned that Spanish variants, Portuguese, indigenous, and Creole languages from her region cannot remain outside global evaluation frameworks .

Bogantes Zamora argues that for AI governance interoperability to be meaningful, evaluation frameworks must be designed to capture the full diversity of the world's languages, cultures, and demographics. This is particularly important for high-risk AI applications in health, justice, and public administration that cannot be responsibly transferred or scaled if countries use incompatible definitions and documentation requirements.

She highlighted that a health system, judicial tool, or public sector algorithm cannot be responsibly transferred, evaluated, or scaled if countries use incompatible definitions of risks, documentation requirements, or audit results . She proposed that before 2027, multilingual evaluation methods should be advanced as part of minimal viable interoperability .

Bogantes Zamora argues that trust in AI cannot be declared through principles alone but must be built through evidence, safeguards, institutional cooperation, and governance frameworks that can operate across borders. She emphasises that adoption moves at the speed of trust, and that this trust must be demonstrated through practical verifiability.

She recalled the point made by her co-chair Rebecca Finlay that 'adoption moves at the speed of trust' , and argued that trust cannot be declared through principles alone but must be built through evidence, safeguards, institutional cooperation, and governance frameworks that can operate across borders .

Singh-Gill frames fragmentation as the opposite of interoperability and identifies four structural challenges it creates. These include regulatory arbitrage where developers move to jurisdictions with weaker oversight, accountability gaps due to global value chains, increased compliance burdens especially for SMEs and low-income country developers, and deepened governance capacity inequality where lower-capacity states are forced to align with dominant frameworks or risk exclusion.

He outlined four specific structural challenges of fragmentation: regulatory arbitrage as developers move to jurisdictions with weaker oversight ; accountability gaps since AI systems operate through global value chains making it difficult to assign responsibility ; increased compliance burdens especially for SMEs, researchers, and developers from low-income countries ; and fragmentation deepening uneven governance capacity and dependency .

Singh-Gill highlights that the scientific panel's report emphasised a shift from static to dynamic system governance, which is also supported by evidence from the AI Governance for Humanity Lab. As agentic AI systems become more autonomous and multi-step, existing governance practices risk being outpaced, requiring more adaptive and anticipatory mechanisms.

He referenced the scientific panel's report presented the previous day, which emphasised the shift from static to dynamic system governance , and noted that work through the AI Governance for Humanity Lab in Valencia also turned up evidence of this shift . He specified that effective governance will need mechanisms such as living risk taxonomies, control testing environments, and staged deployment .

Singh-Gill proposes cross-border regulatory sandboxes as a practical pathway for reinforcing interoperability. By connecting regulatory sandboxes from different environments, countries can test AI applications under regulatory supervision before wide deployment, enabling cross-border test pilots, model evaluations, and evidence sharing without requiring a single governance model.

He described this as one of several 'pathways' for reinforcing interoperability, deliberately called so because they are not prescriptive recommendations but practical guideposts for cooperation . He specifically proposed establishing connections between regulatory sandboxes to enable cross-border test pilots, model evaluations, and evidence sharing .

The EY Representative argues that interoperability does not require all countries to regulate AI in exactly the same way, but can be achieved at the implementation stage through shared technical standards and a focus on functional outcomes. Good governance is not about compliance but about ensuring AI systems operate in the way they were meant to, which is a shared interest of both business and society.

The representative noted that as a global firm operating in virtually every country and supporting both AI adoption and auditing, EY sees strong demand for interoperability between different governance approaches . The representative argued that good governance is not about compliance but about making sure AI systems are really being used in the right way, and that this can be a core basis for interoperability .

Al Hameli argues that full compatibility in interoperability does not need to be the goal, and that a more realistic and effective approach is to focus on selective and deliberate compatibility of the components that matter most for most nations. She grounds this in three principles the UAE uses: trust between partner nations as a precondition, credibility through verifiable accountability, and sovereignty respecting different starting points.

She stated that 'full compatibility in our interoperability... doesn't need to be the goal' and proposed working on 'selective and deliberate compatibility of the components that matter the most for most nations' as a good starting point . She outlined three UAE principles: trust between partner nations as a precondition for any shared framework , credibility ensured by verifiable accountability , and sovereignty respecting that states have different starting points and paths .

Al Hameli identifies the absence of shared incident reporting infrastructure as a critical gap in current AI governance. Different incident monitors track failures independently and voluntarily, with no legal duty to report or recognition between them, meaning that safety failures in one country have no guaranteed path to reach regulators elsewhere before the same failure recurs.

She noted that governance frameworks are multiplying faster than the evidence base underneath them, and that there is no shared incident reporting infrastructure, with different incident monitors tracking failures independently and voluntarily . She highlighted that 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 .

Syed Ahmed argues that while existing standards like ISO 22982 or the EU AI Act can serve as starting points for interoperability, the real challenge lies in building adapters that cover the entire breadth of regulations. High-level intent may be similar across frameworks, but the specific controls required differ significantly, requiring mapping two to three levels deep beyond surface-level alignment.

He used the example of 'transparency' to illustrate the challenge: in ISO, transparency means a model needs to be explainable, while in the EU AI Act it means providing sufficient evidence and system logs - the same intent but not a one-to-one mapping . He described Infosys's attempt to use AI to build an adapter mapping standards, which failed because the mapping needed to go two to three steps down to align individual controls .

Syed Ahmed argues against building AI governance into AI use cases from the start, because governance must be dynamic and cannot assume the same rules will hold over time. Instead, he proposes a separation of AI governance from AI use cases, with a central repository of policies that can be applied and enforced at runtime.

He argued that governance has become more dynamic and cannot be static, and that rules set at the start of a use case cannot be assumed to hold even one month later given how fast things are changing . He proposed building a central repository of policies that can be applied and enforced at runtime, separating AI governance from AI use cases effectively .

Ahmed argues that a shared taxonomy alone is insufficient for interoperability and that what is needed is a meta-model for controls mapping at the lowest level, built into a knowledge graph that can serve as a foundation for interoperability models. The compliance burden on globally operating companies is immense because regulations interact not only with each other but also with other tech regulations across geographies.

He described the compliance burden on a company like Infosys that operates globally, noting that AI regulations must be interoperable with other tech regulations like GDPR and DORA across multiple geographies . He proposed that the committee should invest in building a meta-model for controls mapping at the lowest level and a knowledge graph that can serve as a foundation for interoperability models .

Finlay argues that the independent scientific panel and the global dialogue are complementary and designed to work together from the outset. The panel provides the evidence base while the dialogue provides the direction, and together they form a shared mission for safe, secure, and trustworthy AI.

She referenced Secretary General Guterres's emphasis on common baselines for frontier systems and common methods to verify them , and Maria Ressa's grounding of her address in real-world examples positioning scientific evidence as the antidote to both fear-mongering and techno-utopianism . She stated clearly that 'the panel provides the evidence. The dialogue provides the direction. Evidence and deliberation designed to work together from the outset' .

Finlay proposes three concrete actions: strengthening the independent scientific evidence base, opening up the science through transparency, and advancing progress in the public interest. She argues that transparency does not slow progress but makes it verifiable and credible, and that disclosure alone is insufficient without a system of benchmarks, standards, and third-party evaluation to make sense of it.

She recommended investing in and connecting the work of the scientific panel, driving coherence by connecting to other state of safety reports including from the UK, the Singapore consensus, and the forthcoming Global South Safety Report . She argued that disclosure alone is not enough and that without benchmarks, standards, third-party evaluation, and verification, individual disclosures cannot be turned into a field-wide evidence base that can be trusted .

Finlay argues that closing the gap between innovation and accountability requires asking fundamentally different questions than those currently driving the field. Instead of asking how fast a system can be deployed or how powerful it is, the focus should be on who it serves, who is accountable when it fails, and how verifiable its safety is.

She proposed reframing the questions driving the field: instead of 'how fast can we deploy this', asking 'who does it serve, who is accountable when it fails, and how do we know?'; and instead of 'how powerful is this system', asking 'how verifiable is this system's safety and against what evidence?' . She stated that 'AI systems must be built inclusively, governed accountably, and operated fairly, transparently, reliably, and securely' .

Finlay argues that a shared minimum baseline already exists to a large extent through global commitments, international law, the dialogue, the Hiroshima process, the G7, and pledges from AI summits. What is needed now is to consolidate, share, and further develop these so they are implementable and adaptable to regional contexts.

She stated that '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' . She argued that what is needed is to consolidate, share, and further develop them so that they are implementable and adaptable to regional context .

Finlay argues that multi-stakeholder participation is essential for the legitimacy and success of AI governance efforts, not merely a desirable feature. She emphasises that meaningful participation must be sustained and further supported, including addressing practical barriers such as visa processes that prevent individuals from global communities from attending.

She stated that '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' . She specifically mentioned the importance of ensuring visa processes so that individuals can participate from global communities , referencing the earlier point made by the Center for Responsible AI representative about a colleague being unable to attend due to visa difficulties.

Finlay announces the Global AI Progress Hub as a concrete initiative by the Partnership on AI to advance progress in the public interest. This public platform allows organisations across the responsible AI ecosystem to share their progress against a clear framework, helping to close the gap between innovation and accountability.

She announced that the Partnership on AI had announced the Global AI Progress Hub that week, describing it as 'a public platform where organizations across the responsible AI ecosystem can share their progress against a very clear framework that serves the public interest' . She argued that no single country, government, nonprofit, or research institution can determine on its own what safe, secure, and trustworthy AI looks like .

Iida argues that the OECD AI Principles serve as a foundational instrument for interoperability because they are high-level value-based principles that have already been used as the basis for domestic guidelines and national AI law in Japan. The Hiroshima AI Process builds on these principles to provide a voluntary framework for AI organisations to assess and mitigate risks and disclose relevant information.

He noted that Japan formulated its domestic guidelines and national AI law based on OECD AI Principles , and that the Hiroshima AI Process provides a code of conduct for AI organisations including developers, offering a voluntary framework for risk assessment, mitigation, and disclosure . He described these frameworks as 'a very ambitious multi-stakeholder approach for collaborative governance' and a 'bottom-up approach for the governance' .

Iida reports on practical cross-mapping work between Japan's HITE framework and the ASEAN framework for advanced AI models, which found more commonalities than gaps. Where gaps were found, many could be explained by differences in social or cultural aspects across individual jurisdictions, suggesting that similar cross-mapping work could be applied to other combinations of national or regional frameworks.

He described ongoing work with Singapore and ASEAN countries on cross-mapping between the HITE framework and the ASEAN framework for advanced AI models . The cross-mapping found some gaps but a much bigger number of commonalities, and in most cases where gaps were found, many could be explained by differences in social or cultural aspects from individual jurisdictions .

Demirkoz identifies an 'evidence dilemma' where scientific evidence cannot catch up with AI's rapid progress, as AI capacity doubles every few months. By the time evidence of real harm or capability is compiled, technology has already passed beyond current governance models. Additionally, the 40 different governance models and ethical guidelines documented are highly fragmented and rarely tested, and when tested, it is by the same companies developing the systems.

She stated 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' . She noted that the panel documented 40 different governance models and ethical guidelines, finding them highly fragmented and rarely tested, and when tested, tested by the same companies developing them .

Demirkoz argues that AI evaluation should not focus solely on models but should adopt a broader approach that looks at the interoperability of tools, environments, and users as well. The lack of unified independent evidence outside of the companies developing these models undermines public trust and objective accountability.

She stated that 'we should shift from evaluating just models to evaluating systems' and that because there is a lack of unified independent evidence outside of the companies developing these models, this undermines public trust and objective accountability .

Demirkoz argues that developing nations bear the greatest cost of AI governance fragmentation because frontier models and computing power are concentrated in just two countries. The Global South risks being locked out entirely because they lack localised infrastructure and a common global evidence base they can use.

She stated 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' . She noted that the Global South risks being locked out completely because they don't have localised infrastructure and a common global evidence base .

Demirkoz argues that regions in the Global South are reliant on technologies for which they lack domestic capabilities, meaning they cannot inspect, audit, or culturally tailor AI models to their own needs. This dependency creates a fundamental governance challenge as these regions cannot meaningfully participate in shaping the AI systems that affect them.

She stated that these regions 'are reliant on technologies which they lack domestic capabilities of, so they cannot inspect, audit, or culturally tailor these models to their own needs' . She also highlighted that vulnerable populations such as children and women, as well as civil society and democratic frameworks, are being damaged by AI harms including AI-generated child sexual abuse, deepfake-enabled sexual violence, and automated disinformation .

Demirkoz highlights that beyond individual harms, AI is causing systemic damage to civil society and democratic frameworks through the erosion of shared reality and the undermining of democratic legitimacy. Automated disinformation and AI-generated content are key drivers of this damage.

She stated that 'civil society and democracies, 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' . She also referenced the Secretary General's point about mental health impacts from sycophantic chatbot behaviours .

Cervera Narvas argues that AI governance should follow a principle-based risk approach where higher risk requires higher care and supervision. He uses the analogy of fire — humanity's first transformative technology — to illustrate that even the most powerful technologies can be governed safely through simple but consistently applied rules.

He used the analogy of controlled fire as the first transformative technology of humankind, noting that ancestors developed two simple rules: always use fire with care, and never leave fire unsupervised . He argued that exactly the same rules should apply to AI, and that the EU AI Act has established a principle-based rule where 'the higher the risk, the higher the care and the supervision' .

The UPU Representative argues that in the postal supply chain, one stakeholder's automated decision immediately affects others down the chain, making accountability and oversight critical. The scientific panel's caution that there is no guarantee autonomous agents will follow their instructions means that accountability and oversight must be built by design, not reconstructed from failure.

The representative noted that in the postal supply chain, 'one stakeholder's automated decision, a customs flag, a routine choice, a risk score, immediately affects others down the chain' and that what the report describes as systemic risk is 'an everyday operational reality' for the UPU . The representative cited the panel's caution that 'there is no scientific guarantee that autonomous agents will follow their instructions and that evidence of them departing is already accumulating' .

The Latvia Representative argues that the gap between dominant and underrepresented languages in AI models is widening rather than narrowing, which has both cultural and practical consequences. For users of small languages, AI costs several times more tokens, creating an economic disadvantage on top of the cultural loss.

The representative stated that 'evidence confirms that gap between dominant and underrepresented languages in leading AI models are widening, not narrowing' and that this 'strips away cultural heritage and context, but it also costs several times more tokens when using AI in small languages' . As a positive example, the representative cited an EU-supported open-weight model that achieved equality across more than 30 EU languages .

The Latvia Representative proposes three 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 evaluation as an equal partner.

The representative proposed three specific 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 .

The Latvia Representative highlights the threat of poisoned training data as a concrete information integrity challenge, noting that coordinated disinformation networks can infiltrate global AI models through training data. Despite this risk, there is currently no shared standard for training data origin and no channel to warn others when such contamination is discovered.

The representative described gathering data for an open EU multilingual AI model and having to filter out millions of articles from a coordinated disinformation network infiltrating global AI models . The representative noted that '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' .

Sam Gregory argues that AI is already causing widespread harm through the erosion of shared reality, which he identifies as a systemic safety risk rather than merely an ethics issue. Through Witness's work with frontline civil society and media, he observes synthetic clones impersonating public figures, faked scenes from conflict zones, non-consensual sexual imagery targeting women and girls, and the 'liar's dividend' where real events are dismissed as AI-generated.

He described Witness's global deepfakes rapid response force and work with human rights defenders and journalists worldwide, observing 'synthetic clones, impersonating public figures, faked scenes from conflict zones and relentless non-consensual sexual imagery targeting women and girls' and 'the liar's dividend the panel names real events waved away with the mere claim of ai' . He cited the independent scientific panel's conclusion that 'AI is eroding our shared sense of reality' .

Gregory argues that information integrity must be treated as a systemic cross-border risk and a core focus for interoperability, not just an ethical concern. When people can no longer tell what is real, every other safeguard built rests on sand, making a trustworthy information environment a prerequisite for trustworthy AI.

He stated that 'trustworthy AI is impossible without a trustworthy information environment and that erosion is a systemic safety risk not just an ethics issue' . He argued that 'when people can no longer tell what is real every other safeguard we build rests on sand' and that this deserves attention as the systemic cross-border risk it is .

Gregory argues that meaningful and interoperable authenticity infrastructure and content provenance standards are needed so that people can understand the composition of content they consume. However, standards like the C2PA specification must be built with privacy and human rights at their centre to prevent them from becoming surveillance infrastructure.

He called for 'meaningful and interoperable authenticity infrastructure and constant provenance, so we can all read the recipe of the content and communication we consume' . He specifically noted that 'standards like the C2PA specification must be built with privacy and human rights at their center so that they cannot become surveillance infrastructure' .

Gregory argues that rather than creating new frameworks, the focus should be on making existing frameworks genuinely interoperable. He calls for convergence around existing human rights conventions, international frameworks, and national and regional legislation, with momentum already existing across jurisdictions and UN bodies.

He called for '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' . He noted that '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' and that 'the task is to make it genuinely interoperable' .

Gregory highlights that detection capabilities for deceptive AI are not equitably distributed, with the technology working least well for those most at risk. This creates a compounding vulnerability where the populations most targeted by AI-generated harms are also the least protected by detection tools.

He cited Witness's benchmark showing that 'detection currently works least well for those most at risk, particularly in the global majority' . He called for detection capabilities for deceptive AI that work globally alongside interoperable authenticity infrastructure .

The Brazil Representative argues that the United Nations has a central role in AI governance that goes beyond facilitating dialogue. The UN can help avoid fragmentation, identify convergence, connect existing initiatives, and crucially ensure that developing countries participate not merely as rule-takers but as full co-participants in shaping the rules of AI governance.

The representative stated that '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 governance' . The representative also emphasised that digital sovereignty is not digital isolation but the ability of states to participate in the global digital economy while preserving regulatory autonomy .

The IEC Representative argues that international standards play a distinct and essential role in AI governance by forming the link between different normative layers. They translate principles set out in international law and soft law into practical tools, give governments a basis for national regulation, and give industry a common language to build safe and interoperable products.

The representative described five normative layers of AI governance: international law, international soft law, national law, international standards, and the normative power of what is technically feasible . The representative stated that international standards '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 interpretable products' .

The Pakistan Delegate argues that safety must be defined inclusively because current 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 a purely technical matter but is contextual, shaped by language, culture, and local realities.

The delegate stated that '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' . The delegate argued that 'safety is not purely technical matter. It is a contextual shape by language, culture, and local realities' and called for safety norms to be developed through inclusive multilateral processes under the United Nations .

The Pakistan Delegate argues that trust requires the capacity to verify, and that a nation cannot trust what it cannot assess. The ability to test, evaluate, and assure AI systems remains concentrated where the technology is built, and without deliberate effort, developing countries will be asked to accept safety assurances they have no means to verify.

The delegate stated that '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' . The delegate called for treating 'the capacity to evaluate and govern AI as an integral part of the development of AI to safety itself supported through shared tools, pooled expertise, and cooperation across regions' .

The Center for Responsible AI representative highlights a concrete and immediate barrier to inclusive AI governance participation: visa processes that prevent representatives from developing countries from attending international forums. This is illustrated by the fact that the representative's colleague from India could not attend due to being unable to obtain a visa at short notice.

The representative noted that 'the very fact that I have to present and not Geeta 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 a 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' .

The Center for Responsible AI argues for a structured AI incident reporting framework as a core pillar of global AI governance. This framework must be democratised at the grassroots level, ensure no individual or community is left behind, establish shared responsibilities throughout the AI value chain, and generate real-world evidence of harm to enable proactive rather than reactive governance.

The representative described the Center's work on developing an AI incident reporting framework that enables 'systematic capture, classification, analysis, response, and where possible, resolution of real-world harms alongside clearly defined institutional responsibilities' . The representative argued that such a framework 'provides a strong foundation for AI development and generates real-world evidence of harm, enabling the AI industry and governments to develop feedback loops for better system design' and enables 'proactive learning-oriented approaches that can keep pace with rapidly evolving AI technologies' .

The South Africa Representative argues that global AI governance frameworks must reflect not only the priorities of advanced economies but also the realities and needs of developing countries. This requires that frameworks be inclusive in their design, equitable in their application, and supportive of technology transfer and capacity building, otherwise governance risks reinforcing existing inequalities.

The representative stated that 'global frameworks reflect not only the priorities of advanced economies, but also the realities and needs to developing countries' and that achieving this 'requires that global governance frameworks are inclusive in their design, equitable in their application, and supportive of technology transfer and capacity building' . The representative warned that otherwise 'governance reinforcing existing inequalities' .

The Concordia Representative argues that the rise of agentic AI represents a fundamental shift in governance challenges, as AI agents can now act directly upon the real world both virtually and physically. Because their failures can cause severe harm before a human can intervene, safety standards must strengthen as autonomy increases, similar to how standards for autonomous driving increase from level 0 to level 5.

The representative stated that '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' . The representative drew the analogy of autonomous driving standards from level 0 to level 5, arguing that 'as AI agents become increasingly autonomous in the coming years, the stronger the safety standards must be' .

The Concordia Representative argues that the cross-border nature of agentic AI failures makes international cooperation on safety standards urgent. A malfunction in one country can disrupt shared infrastructure, supply chains, or public safety elsewhere, meaning that true international cooperation requires developing safety standards and red lines inclusively.

The representative stated that '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' . The representative cited the Singapore Consensus from May, where 100 experts from around the world identified 10 foundational principles for managing agentic AI risk .

The Dubai Cable representative argues that as AI agents are authorised to act, the line between a policy document and a production incident becomes very thin. Accountability must therefore be engineered before deployment, with permissions, oversight, and audit trails designed in from the start rather than retrofitted after an incident occurs.

The representative stated that 'once an AI agent is authorized to act, the line between a policy document and the production incident becomes very thin' . The representative argued that 'accountability must be engineered before deployment with the permission, oversight, audit trial need to be designed to build a good governance strategy. A time-based, not added after an incident' .

The Concordia Representative (attributed in the arguments list to Ravin Thambapillai) argues that the model context protocol's rapid adoption demonstrates that open-source technical standards designed around real-world pain points can achieve global de facto status quickly. This shows that governance standards can be developed and adopted rapidly when they meet three criteria: addressing real-world implementation pain points, being open source, and accelerating both AI benefits and governance frameworks.

Thambapillai argues that the model context protocol's rapid adoption from announcement to global de facto standard in 18 months demonstrates that open-source technical standards designed around real-world pain points can achieve rapid global adoption. This shows that governance standards can be developed and adopted quickly when they address real implementation challenges, are open source, and accelerate both AI benefits and governance frameworks.

He described how the model context protocol went from an idea to the global de facto standard adopted by every large AI research lab, every enterprise, and every AI company in just 18 months . He noted that the spec was 'designed around the pain points that people at the frontier doing the implementation of these systems were experiencing' and that it made significant progress on authorization challenges and human oversight in AI tool chains .

Qinghua Lu argues that countries need to agree on a common set of risk management principles and share evaluation methods and results across borders. Given the shift from static to dynamic governance, benchmarking alone is insufficient, and post-deployment evidence from real-world use is essential, particularly for agentic systems that receive human instructions at runtime.

She proposed that countries agree on common risk management principles such as evaluating AI systems through the entire lifecycle, requiring independent third-party participation in evaluation, and not allowing deployment if risk exceeds government-set thresholds . She noted that 'benchmarking itself can provide an initial understanding about the AI system's performance under representative configuration, but it's not sufficient' especially for agentic systems that receive human instructions at runtime .

Lu proposes two complementary approaches to international collaboration on AI governance. Horizontal collaboration focuses on mapping and connecting existing tools and standards through common platforms and dedicated working groups. Vertical collaboration focuses on joint technical work on specific challenges, with the international network of AI safety institutes' joint multilingual capability testing as a concrete example.

She described horizontal collaboration as focused on mapping or connecting existing tools or standards through common platforms, with dedicated working groups building connectors such as mappings between different risk management frameworks or incident taxonomies . She described vertical collaboration as focused on specific technical challenges, citing the international network of AI safety institutes' joint testing exercise on large language models' multilingual capability testing using common methodologies where each country focused on different languages .

Buolamwini argues that AI evaluation frameworks must avoid misleading measures of success, particularly benchmarks that do not represent the global majority. Drawing on her foundational research on algorithmic auditing, she shows that gold standards from major institutions often failed to represent the global south and people with dark skin, meaning apparent progress was illusory.

She described her foundational research on algorithmic auditing, noting that gold standards from institutions like the National Institute for Standards and Technology 'oftentimes did not represent the global south, did not represent people with dark skin' . She stated that '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' .

Malhotra argues that AI must be tailored to local context and that countries should be empowered to harness AI in ways that support their own development goals. The UK's AI for Development Programme exemplifies this approach by building AI tools in more than 40 African languages and supporting locally led AI ecosystems.

She described the UK's AI for Development Programme as working to build AI tools in more than 40 African languages, supporting 13 AI labs, and helping countries develop locally led AI ecosystems that reflect their own priorities . She also described the UK's AI Security Institute as helping build scientific foundations for trustworthy AI through evaluation methodologies, testing, safeguards, and risk assessment, and as part of a growing network of AI safety and security institutes collaborating on evaluation approaches .

Saulo argues that the World Meteorological Organization's 150 years of experience in exchanging data across 193 countries through standards and trust offers directly applicable lessons for AI governance. Key lessons include the importance of verification — checking forecasts against reality — and the centrality of equitable access as the main concern of national meteorological services.

She noted that WMO has been exchanging data across borders for more than 150 years and that 'to exchange data that is usable and by different communities across the world, 193 countries actually, you need to have standards. And you need to have trust also on that' . She described WMO's verification processes where forecasts are checked against reality to set standards for the quality of AI-driven systems , and cited a recent survey of WMO members where 'equitable access' was identified as their main concern and opportunity regarding AI .

The Global Cities Hub representative argues that cities and regions are not merely stakeholders but public authorities with significant responsibilities in sectors where AI is already transforming societies, including healthcare, education, mobility, and public administration. They represent a still-missing governance layer that could contribute to horizontal facilitation of AI governance.

The representative stated 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' . The representative noted that 'AI impacts are experienced first and most visibly in cities by two-thirds of the global population' and that local and regional governments 'represent a still-missing governance layer which could contribute to the horizontal facilitation' .

The UN Resident Coordinator Office representative argues that inclusive governance must ensure all countries have the capacity and voice to shape how AI is developed, deployed, and regulated. UN Resident Coordinators, present in 162 countries and territories, are positioned to accompany member states in translating global AI governance principles into national action.

The representative spoke on behalf of five resident coordinators serving Angola, Eswatini, Costa Rica, Montenegro, and Bosnia and Herzegovina , and cited examples of AI readiness assessments in Montenegro, national dialogues in Eswatini, and public sector AI initiatives in Bosnia and Herzegovina . The representative stated that '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' .

The Ethiopia delegate argues that interoperability and compatibility of AI governance approaches cannot be achieved through isolated national efforts but require continuous dialogue, mutual trust, and strong multilateral cooperation. Ethiopia's commitment to this vision is rooted in a long-term belief in multilateralism and inclusive dialogue under the United Nations.

The delegate stated that 'interoperability and compatibility approaches cannot be achieved through isolated national efforts. They require continuous dialogue, mutual trust and strong multilateral cooperation' . The delegate described Ethiopia's own efforts including an artificial intelligence institute and a planned AI university , and noted that the African Union has designated Ethiopia's Prime Minister as champion for AI and digital health .

The Japan Delegate argues that the Hiroshima AI Process represents a valuable best practice for inclusive international AI governance, having advanced concrete initiatives including guiding principles and a Code of Conduct as the first international norms for generative AI. With approximately 70 countries and regions including the Global South as Friends Group members, it demonstrates how inclusive governance can be achieved.

The delegate stated that 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 the Code of Conduct as the first international norms for generative AI' . The delegate noted that 'approximately 70 countries and regions, including the Global South, have endorsed these principles as the Friends Group, and they share knowledge and experiences related to AI governance and policy in order to achieve trustworthy AI' .

The lead writer of the International AI Safety Report argues that frontier AI risks across three categories — misuse, malfunctions, and systemic societal impact — have moved beyond speculative concerns and are supported by a range of evidence from laboratory studies to real-world incidents. These risks cannot be deferred indefinitely as capabilities continue to improve rapidly.

The writer stated that 'those frontier risks cannot be dismissed as speculative concerns. We have a range of evidence from laboratory studies to real-world incidents that now show that those concerns have merit' . The writer described three categories of risk: misuse including potential lowering of barriers for attacks with bioweapons ; malfunctions as AI becomes more agentic and harder to oversee ; and systemic impact including how AI will affect work, information gathering, and governance .

The Lithuania Delegate argues that trust in AI rests on three pillars, the first being that AI must stay human-centred and grounded in democratic values. The human-in-the-loop principle, requiring human oversight, validation, or intervention at critical decision points, must be preserved as a prerequisite for safe and responsible innovation rather than treated as an obstacle.

The delegate stated that 'the 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' . The delegate also cited Lithuania's experience with persistent cyber threats, disinformation campaigns, and hybrid attacks as evidence that 'resilience must be built into technology from the beginning' .

The Bangladesh Delegate argues for a comprehensive governance framework that combines binding legal obligations with ethical standards, while also being innovation-friendly. Regulatory approaches should be risk-based and include regulatory sandboxes to enable startups and innovators in developing countries to build trustworthy AI without undue regulatory hurdles.

The delegate called for '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' . The delegate also called for a comprehensive governance framework that upholds human rights, transparency, accountability, privacy, and fairness while addressing algorithmic bias, misinformation, disinformation, and AI-enabled gender-based violence .

The France Representative argues that the integrity of electoral processes is threatened by disinformation practices that often make use of AI, making this a high-stakes issue for international governance. The multi-stakeholder approach for global AI governance dialogue and the integration of science and research are essential responses to this challenge.

The representative stated that 'the issue is also high stakes with regard to the integrity of electoral processes threatened by disinformation practices that often make use of AI' . The representative noted that France tackled this issue during the G7 summit and that 56 enterprises signed up to the principles of the Hiroshima Process on AI, targeting common principles and practices for the governance of advanced AI systems .

The representative argues that navigating the labyrinth of international and local laws, regulations, standards, and principles is complicated for any company or individual developing or deploying AI systems. A new international instrument on AI governance must therefore prioritise interoperability and establish legally binding minimum mandatory standards that are truly applicable in all countries across all regions.

The representative stated that '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' . The representative also emphasised the responsibility of academics and others to raise awareness, educate, and advocate for AI to be designed, developed, deployed, and used with ethics in mind .

Virginia Dignum frames the panel discussion around the idea that AI is a set of choices — technical, institutional, and political — and that interoperability is one of the clearest examples of this. She notes that without shared evidentiary standards and mutually recognised approaches, parallel regimes risk neither protecting people nor allowing accountability to travel with the technology.

She stated that 'without shared evidentiary standards and mutually recognized approaches, we risk parallel regimens that neither protect people nor allow accountability to travel with the technology' . She noted that 'we have seen this before in many different fields where fragmentation creates gaps that arm the least resourced first' .

The Conference Organizer explains the procedural framework for managing floor interventions, specifying that member state representatives and other stakeholder representatives should alternate in speaking order. This approach is designed to ensure that both governmental and non-governmental voices receive equal opportunity to contribute to the discussion.

The Conference Organizer stated that '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' . Speakers were required to have inscribed in advance through the UN Global Dialogue on AI website, and a timer was set to ensure all speakers could be accommodated within the allotted time .

The Australia Delegate argues that new AI governance efforts must be informed by tested approaches already developed through organisations such as ISO, ITU, OHCHR, and UNESCO, as well as through the AI Summit Series, the Hiroshima AI Process, and the Global Partnership on AI. 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 contexts.

The delegate stated that 'we should build on existing international work rather than creating frameworks from scratch' and cited ISO AI standards, ITU, OHCHR, and UNESCO guidance as offering common baselines . The delegate also referenced regional initiatives including the APEC AI Initiative, ASEAN's AI Governance Framework, and the African Union Continental AI Strategy as demonstrating how shared principles can be adapted to different economic, social, and regulatory contexts .

The Australia Delegate argues that effective AI governance requires shared principles that can be implemented flexibly across different contexts, and that governance must be understood as an ongoing process rather than a fixed achievement. This reflects the reality that AI technologies and associated risks are constantly evolving.

The delegate stated that 'responsible AI is not a one-off task, but a continuous process of governance, monitoring, adaptation and improvement as technologies evolve and risks emerge' . The delegate also noted that 'a one-size-fits-all approach to AI governance will not be effective' as countries face different risks, priorities, levels of digital development, and regulatory environments .

The Australia Delegate expresses support for the UN Global Dialogue as a platform for multi-stakeholder collaboration on AI governance challenges. By sharing with each other, participants can strengthen the digital ecosystem, enhance connectivity between economies, and contribute to outcomes that benefit all countries.

The delegate stated that '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' . The delegate argued that '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' .

The Co-Chair explains the procedural framework for managing floor interventions during the session, emphasising that statements will be delivered based on a speaker's list established through advance inscription on the UN Global Dialogue on AI website. This system is designed to facilitate the largest number of speakers possible within the available time.

The Co-Chair stated that 'statements will be delivered on the basis of the speaker's list established through the inscription that was made available on the UN Global Dialogue on AI website' . The Co-Chair noted that 'the e-delegate list is closed and any changes has to be communicated to the secretariat' and that 'microphones will automatically be switched off once your time has expired' to ensure broad participation .

The Facilitator UNICC introduces a structured networking exercise designed to move away from formal presentations and engage all participants directly in identifying their single most important priority on AI interoperability. The exercise requires participants to speak with someone they do not already know, ensuring cross-pollination of perspectives.

The facilitator instructed participants to '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' and discuss '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 that one key priority' . The facilitator emphasised that participants should not cheat by talking to someone they already know, underscoring the importance of diverse exchange .

The Facilitator UNICC introduces a second interactive activity where participants write down their identified priority item on post-it notes to be handed to co-chairs as they leave. This mechanism is designed to ensure that the informal discussions held during the networking exercise are captured and feed into the formal co-chair summary report.

The facilitator asked participants to 'write down that priority item' on post-it notes and stated that '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' . The facilitator connected this activity directly to the formal governance process by noting that 'the co-chairs will take those points into consideration' when providing a summary report back to the dialogue of dialogues .

Rachel Adams frames the second panel discussion as a progression from the first panel's identification of existing principles and standards to the practical question of implementation. She identifies three focal areas: what common foundations are already emerging, what practical steps can be taken now, and how countries and communities with fewer resources can be active shapers of global governance infrastructure.

She stated that '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' . She specified that the panel would explore 'what must actually be shared across borders, evidence, measurement systems, evaluation methodologies, reporting tools, institutional practices, so that safe, secure and trustworthy AI can be governed in ways that are compatible and inclusive' .

Rachel Adams argues that a model of interoperability that merely replicates the assumptions and priorities of powerful jurisdictions would be inadequate and potentially harmful. Genuine interoperability must be designed to accommodate different national contexts and priorities, ensuring that less powerful countries are not simply recipients of governance frameworks developed elsewhere.

She stated that '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' . This framing set the agenda for the entire second panel discussion .

Rachel Adams highlights the critical importance of measurement choices in AI evaluation, framing this as a question of rights and opportunities rather than merely a technical matter. She introduces Dr. Joy Buolamwini's work as evidence that the design of evaluation systems has direct consequences for whose experiences and harms are captured.

She stated that 'Dr. Buolamwini, 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' . She asked what must be ensured in shared evaluation systems and standards 'so that these systems capture bias and discrimination rather than simply making AI' more efficient .