Relativity

Make it operational, not declaratory. First, establish working-level cooperation between key institutions. That means connecting bodies like China's CAC (Cyberspace Administration of China), MIIT, and TC260 with counterparts in the EU, OECD, and national AI Safety Institutes. These actors are already shaping testing, standards, and enforcement. Second, align on specific mechanisms. For example: – shared approaches to model evaluation (e.g. CAICT AI Safety Benchmark 2.0 vs. Western eval frameworks) – interoperability of standards (TC260 AI safety standards vs. ISO/IEC processes) – incident response models (China's emphasis on monitoring, early warning, and emergency response) China is already operationalizing them through mandatory model registration, safety testing, and standards systems. Third, create a structured channel for ongoing expert dialogue. Track 2 efforts (e.g. CISS–Brookings, IDAIS) have already produced outputs like shared glossaries and red-line discussions. The UN Dialogue should institutionalize this, not duplicate it. Fourth, link capacity-building with safety infrastructure. China is actively pushing this through UN initiatives and Global South engagement. That should be integrated into the Dialogue rather than treated separately. Success = a short list of joint technical priorities + named institutions responsible for follow-up.

• Safe, secure and trustworthy AI
• AI capacity-building
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

These four are where real convergence is already emerging. Safe AI is already being operationalized in China through CAC-led regulation (model registration, security assessments) and standards like the "Basic Security Requirements for Generative AI Services." The same problems-evaluation, misuse, reliability-exist globally. Capacity-building is central because China is actively shaping this space through UN resolutions, the "Group of Friends," and bilateral initiatives. Ignoring this would mean ignoring one of the main drivers of global governance alignment. Interoperability is the core issue. China is building a standards-based system (TC260, SAC/TC28/SC42), while the West relies more on risk-based regulatory frameworks (EU AI Act, NIST RMF). The challenge is making these systems compatible enough to cooperate. Transparency and oversight are where both systems converge in practice: safety testing, red teaming, labeling, auditability, and human control requirements are present on both sides-even if implemented differently. This is not about harmonization. It's about making systems legible and comparable.

First, crisis coordination for advanced AI incidents. The current themes address safety and accountability, but not how states and institutions should respond when high-impact failures or misuse occur in real time. This gap is becoming more visible. In China, senior policy language now refers to "monitoring, early warning, and emergency response systems" for AI, while standards and policy discussions are moving toward incident response and risk classification. At the dialogue level, initiatives such as IDAIS and other China-Western Track 2 processes have already raised emergency preparedness, red lines, and coordination among domestic AI safety authorities. The Global Dialogue should explicitly address incident notification, shared warning thresholds, and coordination between safety bodies, regulators, and standards institutions. Second, governance of frontier and agentic AI systems. This issue cuts across safety, transparency, and open models, but is not named directly. In China, TC260 has elevated agent safety and AI safety standards, CAICT's AI Safety Benchmark 2.0 now includes frontier risks such as deception, loss of control, and dangerous misuse, and legal scholars are increasingly discussing systemic and extreme risks in proposed AI law frameworks. These developments closely parallel Western debates on frontier model governance, even if the institutional language differs. Third, reducing barriers created by mutual misperception. A practical obstacle to cooperation is that Chinese and Western governance approaches are often viewed through strategic distrust rather than technical comparability. This can discourage engagement even where institutions are addressing similar problems through different tools-for example, China through CAC regulation, CAICT benchmarking, and TC260 standards, and Western actors through risk frameworks, audits, and safety institutes. The Dialogue should therefore support confidence-building through sustained expert exchange, clearer institutional transparency, and joint technical work. Framed diplomatically, this is not about erasing differences. It is about reducing misunderstanding where shared interests already exist.

In the AI governance and knowledge sector, the main gap is not the absence of activity. It is the lack of coordination across governance systems. On one side, there are major advances. China is moving from high-level principles to implementation through CAC regulation, mandatory model registration and security assessments, TC260 standards, CAICT safety benchmarking, and increasing focus on agent safety, frontier risks, and emergency response. At the international level, it is also investing in AI capacity-building and UN-based governance initiatives. For my sector, this creates a major opportunity: there is now enough institutional and technical substance in both Chinese and Western ecosystems to support structured comparison, translation, and mutual learning. This is especially valuable for curators, researchers, and civil society actors working to make governance developments legible across regions. Track 2 dialogues have already shown that cooperation is possible on practical issues such as glossaries, red lines, testing, and emergency preparedness. The main challenge is fragmentation. Governance tools are developing in parallel, but not in an interoperable way. Chinese institutions and Western regulators often address similar risks through different legal, technical, and institutional mechanisms. This makes comparison difficult, slows joint learning, and increases the risk that mutual mistrust will block cooperation where interests actually overlap. A second challenge is uneven access. Capacity-building still often focuses on access to AI capabilities, while access to safety methods, evaluation practices, and governance expertise remains limited. The biggest opportunity is therefore clear: turn parallel governance into usable cooperation. For my sector, that means building shared understanding between regulatory systems, reducing misperceptions, and helping translate concrete governance practice into a form that others can engage with.

The AI Dialogue can play a practical bridging role: not by trying to unify all governance models, but by making them more legible, comparable, and interoperable. Its main value is that it sits within the UN and brings together governments and other stakeholders in a format explicitly designed to discuss international cooperation, share best practices, and support compatibility across approaches. That makes it a rare forum where Chinese, Western, and Global South actors can engage within the same institutional setting. To be useful, the Dialogue should focus on concrete areas where cooperation is already possible. These include risk terminology, model testing and evaluation, transparency practices, incident reporting, early warning, and emergency response. This matters because governance systems are advancing in parallel. China, for example, is already building a standards- and testing-based governance architecture through actors such as CAC, TC260, and CAICT, while also emphasizing monitoring, early warning, and emergency response at senior policy levels. The Dialogue can also reduce friction created by mutual misperception. Chinese and Western governance communities often address similar risks through different institutional languages and tools. A structured UN forum can help shift attention from political assumptions to comparable governance functions. Its added value should therefore be threefold: convene the right institutions, identify specific areas of technical comparability, and create continuity across annual meetings. If it does that, it can become a platform for confidence-building and for steady progress on issues where no country can act effectively alone.

The AI Dialogue should build on existing initiatives that already connect technical, regulatory, and diplomatic communities. First, it should connect with the Independent International Scientific Panel on AI, whose annual report will be presented at the Dialogue. That creates a direct channel between scientific assessment and multi-stakeholder discussion. Second, it should draw on Track 1.5 and Track 2 dialogues that have already produced practical outputs. Examples include the CISS–Brookings dialogue, which published parallel glossaries of AI risk terms, and IDAIS, which has discussed red lines, safety assurance, and emergency preparedness. These efforts show that China-Western cooperation is already possible at the expert level. Third, it should connect with existing standards and testing ecosystems, including China's TC260, SAC/TC28/SC42, and CAICT benchmarking work, as well as international standard-setting bodies such as ISO, IEC, and ITU. WAIC's 2025 Global AI Governance Action Plan explicitly called for faster development of international standards and for shared platforms for mutually recognized AI safety testing. Fourth, it should link to UN capacity-building initiatives, including the AI capacity-building resolution, the Action Plan for Good and for All, and the Group of Friends on AI Capacity-Building, because governance cooperation will fail if safety and capacity-building are treated separately. The Dialogue's added value is not to replace these mechanisms. It is to connect them, surface overlap, reduce duplication, and turn scattered initiatives into a more coherent international governance process.

Different stakeholders should have clearly defined roles, not just speaking opportunities. Governments should present concrete regulatory updates and implementation lessons, for example on standards, safety evaluations, incident response, and public-sector procurement. Standards bodies and technical institutions should be invited to compare methodologies, including work from ITU's AI for Good process, OECD/GPAI, MLCommons, and national AI Safety/Security Institutes. ITU's 2026 Summit is already framed around skills, standards, and partnerships; GPAI operates as a multistakeholder partnership under the OECD; MLCommons is building AI risk and reliability benchmarks; and the International Network of AI Safety Institutes exists precisely to connect technical expertise across countries. The Dialogue should be structured in three layers. First, a plenary for political direction. Second, four small technical roundtables with pre-circulated questions and named rapporteurs: evaluations and testing, incident response, open models, and capacity-building. Third, a public closing session where rapporteurs present 3–5 practical recommendations each. This would produce usable outputs rather than general statements. The UN resolution already provides for a plenary, a presentation of the Scientific Panel's report, and thematic discussions, so the next step is to make those thematic discussions more operational. Civil society and academia should be tasked with producing short evidence briefs in advance. Industry should be asked to submit implementation case studies, not marketing statements. Underrepresented regions should receive travel support and reserved speaking slots, consistent with the resolution's encouragement of support for developing-country participation. A useful rule would be simple: every session should end with one concrete cooperation proposal, one institution able to carry it forward, and one timeline.

Three groups remain underrepresented. First, regulators and technical institutions from the Global South are still too often invited as policy recipients rather than co-designers. This matters because the UN process is explicitly meant to help close AI divides, and UNESCO's Observatory and progress tracking already show wide variation in countries' readiness and implementation capacity. The Dialogue should reserve seats for regulators, standards officials, and public-interest technologists from low- and middle-income countries, and fund their participation directly. Second, standards and testing communities are underrepresented relative to high-level policy actors. Bodies such as TC260 and CAICT in China, MLCommons, ISO/IEC experts, and AI Safety/Security Institutes are often where governance becomes operational through benchmarks, evaluations, and technical guidance. They should not be peripheral participants. They should be at the center of the Dialogue's technical tracks. Third, Chinese expert and regulatory perspectives are still not integrated consistently enough into global discussions. There is a tendency to discuss China as an object of governance rather than as a source of implementation experience. That is a mistake. China's institutions have built substantial practice in model registration, standards, safety benchmarking, agent safety, and capacity-building diplomacy. The Dialogue should include Chinese regulators, think tanks, and technical evaluators in working-level sessions, especially on interoperability, testing, and incident response. Inclusion should not mean symbolic diversity. It should mean shared drafting roles, funded participation, and guaranteed places in technical working sessions.

Three formats would work especially well. 1. Regulator-to-regulator clinics. Small closed-door sessions where agencies present one live governance problem and peers respond with implementation examples. For example: model evaluations, incident reporting, or governance of agent platforms. This would be more useful than generic panels. 2. Standards and benchmark labs. A hands-on format where institutions such as MLCommons, AI Safety/Security Institutes, and national testing bodies compare evaluation methods, taxonomies, and reporting templates. That would make interoperability concrete rather than rhetorical. Existing initiatives already provide a base: MLCommons develops AI risk benchmarks, the International Network of AI Safety Institutes connects technical expertise, and WAIC's 2025 Action Plan called for shared platforms for mutually recognized AI safety testing. 3. Structured red-line and crisis simulation exercises. Mixed groups of governments, labs, standards bodies, and civil society should work through realistic scenarios: cross-border misuse, model failure in a critical sector, or release of a high-risk open model. China-Western Track 2 dialogues have already explored red lines, glossaries, emergency preparedness, and crisis coordination. The Dialogue should build on that experience in a more institutionalized setting. Two design rules matter. First, every format should use pre-read briefs and defined questions. Second, every session should produce a short written output: points of agreement, disagreement, and next steps.

1. China's CAC registration and security assessment system China's Interim Measures for the Management of Generative AI Services created a practical pre-deployment governance mechanism: model providers must undergo registration and security review before certain public deployment, and by March 2025 the CAC had registered 346 models. That is an operational gatekeeping tool, not just a principle. 2. TC260 and CAICT as implementation infrastructure China's TC260 has developed an AI safety governance framework and standards work on generative AI security, incident response, agent safety, and safety guardrails. CAICT's AI Safety Benchmark 2.0 now covers frontier risks, scenario safety, and agent safety, including deception, loss of control, and dangerous-domain misuse. Together, they show how governance can move from policy to standards and evaluation. 3. MIT AI Risk Repository MIT's AI Risk Repository compiles 1,700+ risks from 74 frameworks, while its mitigation map organizes 800+ mitigations and its incident tracker classifies 1,400+ real-world incidents. This is a strong model for building shared taxonomies and evidence-based governance. 4. OWASP, ISACA, and Cloud Security Alliance OWASP's Top 10 for LLM Applications gives a concrete security baseline for generative AI systems, especially around prompt injection, insecure output handling, and training-data risks. ISACA provides practical governance tools such as its AI Governance Brief and AI Audit Toolkit for enterprise oversight and assurance. The Cloud Security Alliance has issued an AI Model Risk Management Framework and launched an AI Safety Initiative to develop practical guidance for safe, responsible, and compliant AI deployment.