CeSIA - Centre pour la Sécurité de l'IA

The first Dialogue arrives at a moment of growing international consensus on AI risks. CeSIA sees four outcomes as essential to making it count. The first is a shared risk taxonomy for the most dangerous frontier AI risks, grounded in the work of the UN Independent International Scientific Panel and the International AI Safety Report. Both should receive formal mandates and sustainable funding as a Dialogue output. Political conditions are more favourable than they may appear: recent statements from leaders across traditional divides, including explicit acknowledgments from the United States and China that AI capabilities in dangerous hands represent a shared security challenge, signal that the moment for naming these risks clearly has arrived. The second is progress toward binding international frameworks. Voluntary commitments have reached the limits of what they can achieve on their own. The Dialogue is well-placed to initiate negotiations toward a framework convention that establishes core principles and enforcement mechanisms. The third is agreement on specific red lines. Some concern capabilities that should not be developed or deployed regardless of the actor: AI assistance in developing biological, chemical, radiological or nuclear weapons; autonomous offensive cyberweapons; and AI systems that can replicate or modify themselves without authorised human oversight. Others concern prohibited uses: lethal autonomous systems that select and engage targets without meaningful human control; delegation of nuclear command decisions to AI; and mass AI-enabled surveillance without democratic accountability. The fourth concerns the architecture for follow-through: intersessional working groups by September 2026, a mid-cycle ministerial review, and a formal commitment to negotiate a framework convention ahead of any subsequent Dialogue session. Substantive outcomes need a process architecture to sustain them.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Interoperability of governance approaches
• Protection and promotion of human rights

Safe, secure and trustworthy AI is our primary concern. No international mechanism currently defines or enforces minimum safety standards for frontier AI systems. The absence of agreed capability thresholds creates conditions for regulatory arbitrage and severe misuse. Six global workshops by CeSIA and The Future Society, with participants from over twenty countries, surfaced the same six risk domains as priorities: weapons of mass destruction risks, loss of human control over AI systems, offensive cyber capabilities, lethal autonomous weapons, harmful manipulation, and child safety. These overlap with, but are distinct from, the Q8 red lines proposals. A 2025 survey by The Future Society found binding red lines to be the most widely supported governance measure among the 44 civil society organisations, research institutes, and think tanks surveyed. That convergence across very different stakeholders is meaningful. Transparency, accountability, and human oversight are prerequisites for any governance regime that can actually be enforced. Without mandatory incident reporting, independent third-party evaluation, and clear accountability for frontier model developers, no other commitment can be verified. Interoperability is essential because fragmentation creates opportunities for regulatory arbitrage. Divergent standards across the EU AI Act, the G7 Hiroshima Process, voluntary safety commitments, and national frameworks risk allowing inconsistencies to be exploited. Identifying risks that would be unacceptable across jurisdictions, or that would threaten international security and stability, is among the most practical contributions the Dialogue can make. Human rights law and existing arms control commitments provide normative grounding for several of these obligations. States that have ratified the Biological Weapons Convention, the Geneva Conventions, and the International Covenant on Civil and Political Rights have accepted operative principles that extend naturally to dangerous AI capabilities. This matters diplomatically: extending existing commitments to AI is a matter of consistency, not innovation, and gives governance advocates a concrete legal basis to work from.

Three issues cut across all the thematic areas. The first is the concentration of AI capacity. By most estimates, the US accounts for roughly 75% of global frontier AI compute, China for around 15%, and the rest of the world for 10%. Building sovereign AI capacity cannot be achieved through technology access or deployment support alone. The realistic path is collective action: mid-sized economies, which the Blueprint for Multinational Advanced AI Development, co-authored by CeSIA and Oxford Martin AIGI and endorsed by Yoshua Bengio, terms 'bridge powers', pooling compute, talent, and data to participate in AI development, not merely its deployment. The Dialogue can help legitimise and coordinate such coalitions. The second is emergency response. No international framework exists for coordinated response to a serious AI-related incident, whether an AI-enabled cyberattack, an AI-assisted bioterrorism event, or a critical infrastructure failure. The United States has begun addressing this at the national level through its AI Action Plan (2025), which called for AI-specific incident response frameworks. The Dialogue can help build the international architecture that national-level efforts increasingly point toward. The third is the transition from principle-sharing to rule-setting. The Dialogue's mandate emphasises exchange of best practices, and this is valuable. The Dialogue is also well-placed to begin the transition toward enforceable obligations, a step that the nature and pace of AI development makes increasingly important.

Current governance gaps are already producing measurable effects. Peer-reviewed research has found AI models generating detailed virology protocols that independent assessors rated at or above expert level (Mouton et al., Carnegie Mellon, 2025). At the same time, threat intelligence reports consistently document AI tools being used to accelerate ransomware campaigns. These are early signals of the capability misuse that adequate governance is designed to prevent. At the European level, the EU AI Act represents real progress in risk-tiered regulation. Its GPAI provisions apply fines of up to 3% of global annual turnover for non-compliant frontier model providers, which is significant. Several major AI providers have not signed the voluntary Code of Practice, however, and fines at this level may not deter the largest developers. The risks these systems pose also do not stop at EU borders: a dangerous model deployed anywhere can create harms everywhere.

The Dialogue has something no other AI governance forum possesses: universal political legitimacy. The OECD, G7, and EU AI Act have each generated important substantive work within their respective scopes. Only the Dialogue brings all Member States together, with the authority that universal participation provides. A common objection is that geopolitical competition makes international AI agreements unworkable. The evidence runs the other way. At the 2024 APEC summit, the United States and China agreed that humans, not AI, must retain control over nuclear weapons, the first bilateral political understanding of its kind between the two leading AI powers. In February 2025, 27 states signed the Paris Declaration on Maintaining Human Control in AI-Enabled Weapon Systems, committing not to delegate life-and-death decisions to autonomous systems operating without human oversight. Cold War precedents, including the Biological Weapons and Chemical Weapons Conventions, show that states can agree to prohibit the most dangerous capabilities even amid deep disagreement. AI red lines are feasible for the same reason. Among the steps that CeSIA believes would be most consequential, and that fall within or near the Dialogue's mandate: formalising the Independent International Scientific Panel with sustainable funding and a mandate covering catastrophic and systemic risks; creating structured mechanisms for non-frontier states to shape governance priorities collectively; opening a dedicated negotiation track toward a framework convention on AI safety, comparable in architecture to the Chemical Weapons Convention; and authorising an international auditing function, modelled on the IAEA, to monitor compliance with agreed frontier AI safety standards. The Dialogue's distinctive contribution is scope: it can connect existing initiatives, give political weight to expert consensus, and provide the universal membership that regional or club governance cannot. That combination is what makes binding AI safety frameworks achievable at scale.

Several existing initiatives provide important foundations. The UN Independent International Scientific Panel on AI is among the most important resources the Dialogue has. It should be formally institutionalised with sustainable funding, and its reports should drive policy outputs rather than informing discussions without follow-through. The AI Summit Series has produced voluntary safety commitments and the International Network of AI Safety Institutes (INASI). The Dialogue should create verification mechanisms for those commitments and a pathway toward binding standards, rather than duplicating the Summit's convening function. The OECD's Hiroshima AI Process (HAIP) Reporting Framework, launched in February 2025, is the first international tool for organisations to disclose their AI governance practices against a common standard. The Dialogue should extend this transparency infrastructure to the full UN membership. The Global Call for AI Red Lines, co-led by CeSIA, The Future Society, and CHAI, endorsed by 15 Nobel laureates and Turing Award winners including Nobel Peace Prize laureate Maria Ressa and Yoshua Bengio, demonstrates that political appetite for binding prohibitions on the most dangerous AI capabilities already exists across geopolitical lines. The Dialogue should build on that momentum. The Singapore Consensus on AI Safety Research Priorities identifies the development of technical risk thresholds as an urgent research need. The Dialogue should connect this scientific agenda to its governance outputs to ensure expert consensus on capability thresholds informs standard-setting. Each of these initiatives operates within a bounded scope: regional, thematic, or club-based. The Dialogue's value lies in bringing them under a single universal roof, providing the political authority that can translate existing momentum into binding commitments.

Civil society and academia occupy a distinctive position in AI governance that neither governments nor industry can fully replicate. Civil society offers a distinctive perspective on long-term public interest that complements government and industry voices. CSOs operating without commercial or electoral pressures can raise concerns that other stakeholders may be reluctant to voice, and bring expertise in translating technical risk assessments into governance demands. CeSIA's bilateral diplomatic engagement across foreign ministries and multilateral forums shows that civil society can carry substantive diplomatic weight when given appropriate access. Meaningful participation would be supported by speaking opportunities in thematic sessions, access to working documents in advance, and channels through which civil society can raise concerns with the Dialogue's organisers. Technical and scientific experts should be integrated into thematic sessions through structured policy-translation work rather than general scene-setting presentations. Bridging the gap between scientific understanding of AI risks and actionable policy responses is one of the most important things the Dialogue's format can enable. Industry participants can make a substantive contribution by bringing current safety practices, evaluation approaches, and incident handling mechanisms to the table, building on the Frontier AI Safety Commitments from the 2024 Seoul Summit. Other stakeholders, including non-frontier governments, regional bodies, and affected communities, each bring perspectives that frontier actors tend to underweight, and the Dialogue's structure should ensure their voices shape substantive outcomes.

The current AI governance conversation is dominated by a small number of frontier AI states and large technology companies. Several categories of perspective remain significantly underrepresented. Civil society organisations focused on catastrophic and long-term AI risks are consistently underrepresented in intergovernmental forums, where near-term harms tend to dominate the agenda. The distinction matters practically: addressing catastrophic risks requires binding prohibitions, red lines, and emergency response protocols, while near-term harm prevention requires different regulatory tools. The Dialogue needs dedicated space for both. Global South governments and civil society face compounding disadvantages: limited technical capacity to assess frontier AI risks independently, limited leverage in informal AI governance spaces, and governance frameworks built primarily around the concerns of advanced economies. Dedicated capacity-building for AI governance expertise is essential, distinct from AI adoption support. These voices should shape the agenda, not simply receive its outputs. Independent technical experts need stronger institutional standing in intergovernmental forums. The UN Independent International Scientific Panel is a significant step in this direction. The Dialogue should build on its mandate, ensure its findings are systematically integrated into policy outputs, and create complementary mechanisms for independent researchers from academia and civil society to contribute alongside it. Mid-sized economies in Europe, Asia, and Latin America, those neither at the frontier nor in the Global South, face a distinctive vulnerability and have a specific interest in multinational AI partnerships. Their collective perspective on sovereignty and dependency deserves structured representation. Practical measures to address this: dedicated participation pathways for Global South civil society, including reserved slots and travel support; funding for translation and remote access; preparatory consultations that reach beyond Geneva-based organisations; and independent civil society rapporteurs producing public assessments of Dialogue outcomes.

Several engagement formats would make the Dialogue substantially more effective. CeSIA and The Future Society have developed a workshop format, first presented at the IASEAI conference at UNESCO Paris, combining technical briefings on specific risk scenarios with facilitated diplomatic exchange. It has helped government participants move from broad safety commitments toward substantive discussion of specific capability thresholds. This model, or something built on it, would work well for the Dialogue's safety and interoperability sessions. With several hundred submissions expected, the Co-Chairs will need tools to surface convergence. Pol.is, used in Taiwan's vTaiwan consultations, identifies positions with cross-partisan support through statistical clustering. A similar instrument applied to the Dialogue's thematic questions could reveal where consensus already exists, including on AI red lines, without requiring manual synthesis of individual responses. A short structured questionnaire distributed alongside the written consultation would serve the same purpose at lower cost. Science-to-policy workshops, structured around specific policy questions rather than general presentations, can produce actionable outputs. A session on 'what do current AI risk assessments imply for capability thresholds?' generates useful material in a way that broad expert overviews do not. Three further formats are worth adopting: Track 1.5 sessions combining officials and civil society experts without attribution, to allow frank discussion of politically difficult questions; an independent civil society rapporteur process producing public assessments of outcomes between sessions; and online participation with real-time translation to extend access beyond Geneva.

The Global Call for AI Red Lines, co-led by CeSIA, The Future Society, and the Center for Human-Compatible AI, demonstrates that specific, enforceable prohibitions on dangerous AI capabilities can attract broad support across geopolitical lines. Endorsed by 15 Nobel laureates and Turing Award winners, launched at the 80th UN General Assembly by Nobel Peace Prize laureate Maria Ressa and presented at the UN Security Council by Yoshua Bengio, it shows that the political constituency for binding international standards is already emerging. The prohibitions it proposes are concrete enough to form the basis of treaty language. Compute security offers a governance model that most international frameworks have not yet engaged with. Advanced AI chips are made by a small number of companies in a small number of jurisdictions. Treating them as trackable, regulated assets, catalogued from fabrication to end use on the model of fissile material accounting, would give oversight bodies a technical lever that behavioural compliance frameworks alone cannot provide. This approach is developed in CeSIA's article 'Why an International Agreement Establishing Red Lines for AI Is Both Necessary and Realistic.' The EU AI Act's General Purpose AI Code of Practice demonstrates what ambitious multi-stakeholder governance can produce. Hundreds of organisations participated in its drafting; most leading AI developers signed on to concrete obligations on transparency, safety testing, and incident reporting. That some chose not to is itself an argument for moving toward binding requirements. An analogous international instrument, convened under the Dialogue's auspices, could generate comparable commitments at a global scale. The EU AI Act's risk-tiered framework, with its prohibited applications and conformity assessment infrastructure, remains the most developed legislative template available for international standard-setting.