Centre for International Cooperation on AI

In my opinion, the first Global Dialogue would succeed if it produced three concrete results. First, a public commitment registry. More than 60 countries have published AI strategies. The EU AI Act, the Bletchley Declaration, the Hiroshima Process and the Global Digital Compact each create separate obligations. No single instrument tracks who committed to what, when, under which framework, and whether those commitments have been met. The Dialogue should mandate the creation of a structured, publicly accessible registry of AI governance commitments across all existing instruments. Without this, every subsequent discussion rests on unverified claims. Second, an honest assessment of verification gaps. The Independent International Scientific Panel on AI will present its inaugural report at the Dialogue. The Panel should be asked to address a specific question: which AI governance commitments are currently verifiable, and which are not? Nuclear governance works because reactors are physical objects subject to inspection. AI development is opaque. Until the international community names this gap directly, governance frameworks will remain statements of intent. Third, a structured mechanism for ongoing institutional input between Dialogues. Annual convenings lose momentum without a permanent process for receiving, reviewing and responding to stakeholder contributions. The Dialogue should establish a standing call for written inputs with a defined review cycle, not a one-off consultation. What the Dialogue should avoid is producing another declaration of principles. The Governing AI for Humanity report (2024) identified the core structural problems: fragmented governance, concentrated development, and uneven participation. Resolution 79/325 created the institutional space. The task now is operational: build the infrastructure that makes cooperation verifiable and accountable. (Disclaimer: Text created with the help of Claude Opus)

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight
• Social, economic, ethical, cultural, linguistic and technical implications of AI

Interoperability is the most urgent priority because it is the precondition for all others. More than 60 national AI strategies, the EU AI Act, the Hiroshima Process and the Bletchley Declaration each define terms, risk categories and compliance expectations differently. Without a common framework for comparing and aligning these approaches, governance fragments further with every new instrument adopted. The Global Dialogue is the only existing platform with the mandate and convening power to address this. Safe, secure and trustworthy AI is a priority because the current approach to AI safety relies on voluntary commitments from developers. Voluntary frameworks function only when participation is rational for the actors involved. History shows this: the IAEA succeeded because states gained legitimacy and access to civilian nuclear technology in exchange for accepting inspections. Frontier AI companies have no equivalent incentive to submit to external oversight. The Dialogue should address what institutional arrangements would make safety commitments binding and verifiable. Transparency, accountability and human oversight connects directly to the verification problem. International governance functions when compliance is observable. For AI, the question is what forms of transparency are technically feasible and what oversight mechanisms would give them institutional force. This is not a principled aspiration; it is an engineering and institutional design problem the Dialogue should treat as such. The social, economic, ethical, cultural, linguistic and technical implications of AI matter because governance discussions dominated by safety and security risk are overlooking the structural effects AI is already producing: labour displacement in creative industries, concentration of training data value in a small number of corporations, and the erosion of cultural and linguistic diversity in AI outputs. These are governance questions, not side effects. (Disclaimer: Text created with the help of Claude Opus)

One structural issue is absent from the listed themes: the political economy of AI cooperation. The seven thematic areas describe what needs governing. None addresses why cooperation on these issues is difficult to build and difficult to sustain. AI governance faces four specific obstacles. No global authority holds enforcement power over AI development. AI capabilities are extremely difficult to verify externally; nuclear reactors and chemical weapons stockpiles are physical objects subject to inspection, but AI models are not. States and corporations hold asymmetric information about capabilities and risks. First-mover incentives reward speed over coordination, because any actor exercising restraint while others do not loses strategic advantage. These obstacles have appeared before. The IAEA for nuclear governance, CERN for collaborative research, Gavi for vaccine distribution, Intelsat for satellite communications: each succeeded because participation served the interests of the actors with the most power to defect. The United States created Intelsat in the 1960s despite holding a monopoly on satellite technology, because the multilateral route gave it legitimacy and influence over developing nations it would not have gained by acting alone. The same logic applies to AI. The Dialogue should create space for a specific question: what institutional arrangements make cooperation on AI rational for the states and companies whose participation is necessary? This question is not captured by any single thematic area. It is why interoperability stalls, why safety commitments remain voluntary and why capacity-building efforts lack sustained funding. Treating institutional design as a cross-cutting theme would give the Dialogue a sharper analytical foundation across all seven mandated areas. (Disclaimer: Text created with the help of Claude Opus)

CIC is an independent think tank focused on institutional design for international AI cooperation. The governance gaps in our selected thematic areas affect our work in two specific ways. The first is a missing step in the policy conversation. Governments, international organisations and civil society groups broadly agree on the risks of uncoordinated AI development. Principles exist. Regulatory frameworks are emerging. But no institution is working on the prior question: how to change the incentive structures driving the race toward AGI. Fewer than ten companies, concentrated in the United States and China, are building frontier AI systems under competitive pressures where restraint is penalised, and speed is rewarded. This game-theoretic dynamic is well understood. What does not exist is a credible plan for shifting it toward cooperation. Funders and institutional partners recognise this problem when CIC describes it. They do not see a clear path to solving it, which makes the work harder to fund and harder to scope. The second is the gap between diagnosis and institutional design. Reports such as Governing AI for Humanity (2024) identify the structural problems with precision: fragmented governance, concentrated development, uneven participation and voluntary commitments with no verification. These diagnoses are correct. The missing layer is operational: what institutional arrangements would make cooperation rational for the actors whose participation is necessary? CIC works on this question, but the field as a whole has not yet developed the frameworks, evidence-based or tested models to answer it with confidence. The opportunity is specific to this moment. The Global Dialogue marks a shift from agreeing on principles to asking how cooperation works in practice. This is the space where institutional design research becomes directly relevant to the policy process. (Disclaimer: created with the help of Claude Opus)

The Dialogue is the only platform with a UN General Assembly mandate to convene governments and stakeholders on AI governance. This gives it a function no other initiative holds: the authority to move the international conversation from agreeing on principles to building operational infrastructure for cooperation. Three specific functions matter. The first is making commitments visible. Over 60 countries have published AI strategies. Multiple international instruments create separate obligations. No public registry tracks these commitments, their overlaps or their gaps. The Dialogue is positioned to commission this registry and to ask the Independent Scientific Panel to assess which commitments are verifiable with current methods and which are not. The second is naming the incentive problem directly. International cooperation on AI will not emerge from shared recognition of risks. It will emerge when the leading AI-developing states and companies conclude they gain more from cooperation than from unilateral action. The IAEA, Intelsat and Gavi each succeeded because they offered the dominant players something they valued: legitimacy, market access, influence over global norms. The Dialogue should dedicate substantive agenda time to this question rather than treating it as implicit. The third is creating continuity between annual sessions. A single annual convening loses momentum without a structured process for receiving and reviewing stakeholder inputs throughout the year. The Dialogue should establish a standing mechanism for written contributions with a defined review cycle and published responses. The added value of the Dialogue is its convening authority under Resolution 79/325. If it uses this authority to build tracking, verification and incentive-alignment infrastructure, it becomes the institutional anchor for AI cooperation. If it produces another set of principles, it duplicates work already done elsewhere. (Disclaimer: created with the help of Claude Opus)

The Dialogue should build on five existing mechanisms, each addressing a different layer of the cooperation problem. The Independent International Scientific Panel on AI provides the evidence base. Its mandate to produce annual assessments of AI capabilities and risks mirrors the function the IPCC performs for climate. The Dialogue should treat the Panel's findings as the starting point for policy discussion at each session, not as a standalone report. The OECD AI Policy Observatory maintains the most comprehensive cross-country dataset on AI policies and governance frameworks. The Dialogue should draw on this data to support interoperability analysis rather than building a parallel tracking system. The UNESCO Recommendation on the Ethics of AI (2021), adopted by 194 Member States, is the only near-universal normative instrument on AI. Its Readiness Assessment Methodology offers a tested tool for evaluating national implementation. The Dialogue should reference implementation data from this process when assessing governance gaps. The EU AI Act provides the most advanced binding regulatory framework for AI. Its risk classification system and conformity assessment requirements are reference points for interoperability discussions, particularly as CEN and CENELEC develop the underlying technical standards. The Bletchley, Seoul and Paris summit processes created voluntary commitments from frontier AI developers. These commitments are not monitored. The Dialogue brings added value by asking what a monitoring mechanism would require and whether the Scientific Panel holds the appropriate mandate to perform this function. The Dialogue's unique contribution is institutional. Each initiative listed above operates independently. None has the mandate to compare commitments across instruments, assess their compatibility or track implementation. Resolution 79/325 gives the Dialogue this mandate. The first session in July is the opportunity to put it into practice. (Disclaimer: created with the help of Claude Opus)

Different stakeholders hold different pieces of the cooperation problem. The Dialogue's structure should reflect this. Governments hold the authority to make binding commitments and the data on national implementation. Their contribution is most useful when it is specific: reporting which instruments they have signed, what implementation steps they have taken and where they face obstacles. The Dialogue should ask governments to submit structured implementation updates, not general policy statements. The private sector holds information no other actor has: data on AI system capabilities, training data provenance, safety testing results and deployment decisions. Frontier AI developers should be asked to provide concrete disclosures on these topics as a condition of participation, following a standardised reporting template. Civil society and academia contribute independent analysis, including on questions governments and companies have limited incentive to raise. The Dialogue should create a structured written submission process with a defined review cycle. Submissions should be published and referenced in the Co-Chairs' summary. A one-off call for inputs without a visible feedback loop discourages serious engagement. International organisations bring coordination capacity and existing data. The OECD AI Policy Observatory, UNESCO's Readiness Assessment Methodology and the ITU's AI for Good platform each hold relevant datasets. The Dialogue should commission these bodies to produce interoperability analyses rather than duplicating their work. On structure: the July session should allocate at least one thematic block to institutional design questions, specifically how cooperation mechanisms function and why they fail. Plenary statements are useful for signalling political will. Working sessions focused on specific operational problems produce more actionable outputs. (Disclaimer: created with the help of Claude Opus)

The Governing AI for Humanity report (2024) found that 118 countries participate in none of the seven major non-UN AI governance initiatives. These are predominantly Global South states. They are the most affected by decisions made in Washington, Beijing, Brussels and a small number of corporate headquarters, and the least able to shape those decisions. This is a structural problem for the Dialogue. AI governance frameworks designed without input from these countries will lack legitimacy and will face resistance at the implementation stage. The countries excluded from current discussions are the same countries where AI systems trained on English-language data are being deployed with minimal local oversight, where platform governance decisions made abroad determine what cultural content is visible, and where the economic effects of AI-driven automation arrive without corresponding investment in transition infrastructure. Inclusion requires more than open registration. Three steps would make participation real. First, funded travel and accommodation for delegations from states with no existing AI governance capacity, particularly from Sub-Saharan Africa, Central Asia and Small Island Developing States. Second, preparatory briefing materials were produced in multiple languages and designed for governments engaging with AI governance for the first time. Third, a structured onboarding process connecting new participants with experienced delegations and civil society organisations before the July session. CIC has established networks in the post-Soviet space and the Gulf region. Both areas are underrepresented in current AI governance discussions despite active national AI investment programmes. CIC is positioned to support outreach and engagement with researchers and policymakers in these regions as part of the Dialogue's preparatory process. (Disclaimer: created with the help of Claude Opus)

Three formats could complement the Dialogue's plenary sessions and produce structured outputs between annual sessions. The first is structured foresight exercises. Futures methodology allows participants to work through specific governance problems under different conditions. For the July session, a scenario exercise could ask: What institutional arrangements for AI cooperation function if a major AI-developing state declines to participate in multilateral frameworks? What changes if a major AI incident requires rapid international coordination? Scenario exercises produce concrete policy options because they require participants to respond to specific situations. CIC develops and facilitates foresight workshops on AI governance and would welcome the opportunity to support the design and delivery of scenario exercises for the Dialogue. The second is Delphi-style expert consultations. A structured, iterative survey of experts across regions and sectors would surface areas of genuine consensus and identify where disagreements are substantive rather than terminological. This method is well established in technology assessment and policy foresight. It produces quantifiable outputs and works asynchronously, making it accessible to participants across time zones and languages. Running a Delphi process in advance of the July session would give the Co-Chairs an evidence base for structuring thematic discussions. The third is commitment-mapping workshops. Participants would receive a structured overview of existing AI governance commitments across instruments and work in groups to identify overlaps, conflicts and gaps. The output would be a draft interoperability assessment that the Dialogue builds on at each subsequent session. These formats share a principle: they generate working documents that each session builds on, creating institutional continuity between annual meetings. CIC's expertise in futures methodology and participatory workshop design is directly applicable and available to the Dialogue's preparatory process. (Disclaimer: created with the help of Claude Opus)

Four historical cooperation mechanisms offer tested institutional designs relevant to AI governance. The IAEA (established 1957) solved a problem close to the one AI governance faces now: how to enable peaceful use of a dual-use technology while preventing its weaponisation. The IAEA created an inspection and verification regime states accepted because participation gave them access to civilian nuclear technology. The institutional design principle is specific: tie access to compliance. Gavi, the Vaccine Alliance (established 2000) solved a different problem: how to distribute a critical resource when the private sector has no commercial reason to serve low-income markets. Gavi pooled funding from wealthy governments and foundations, negotiated volume-based pricing with pharmaceutical companies and helped countries build delivery infrastructure. Companies gained a guaranteed market. Countries gained vaccines. The principle: make participation financially rational for all parties. Intelsat (established 1964) demonstrates how a dominant actor chose multilateral cooperation over unilateral action. The United States held a monopoly on satellite technology but created a multilateral organisation because the cooperative route offered legitimacy and influence over developing nations. The principle: cooperation succeeds when it serves the strategic interests of the most powerful participants. CERN (established 1954) shows how shared infrastructure reduces duplication and enables research that no single state could afford alone. Member states pool resources for facilities and share the results. The principle: joint investment in infrastructure lowers the barrier to cooperation. Each mechanism addresses a specific structural obstacle: verification, distribution, incentive alignment, and shared infrastructure. AI governance faces all four obstacles simultaneously. The Dialogue would benefit from a systematic analysis of how these institutional designs translate to AI, where they apply directly and where adaptation is needed. CIC's research programme is focused on this translation work. (Disclaimer: created with the help of Claude Opus)