AI Safety Asia

Beyond shared values and convergence on governance, we believe the Dialogue's first session should establish concrete follow-through architecture that can operate between annual sessions. Success should be measured not only by participation, diplomacy or consensus language, but by whether the Dialogue produces a small number of implementation-ready mechanisms: • A limited and prioritised set of follow-up workstreams, attached to specific themes from Resolution 79/325, with substantive contributions and leadership from the Global Majority, rather than open-ended agendas that lose momentum between sessions. • Named, regionally balanced co-leads or coordination contacts for each thematic stream — whether designated Member States or independent experts, drawing where useful on the special-rapporteur model from international human rights — backed with substantive and functional resources, with clear responsibility for advancing work between annual convenings. • A clear, dual-track reporting and learning cycle that allows Member States and other stakeholders to track progress, identify implementation gaps, and propose corrections, drawing where useful on models such as the UN Human Rights Council's Universal Periodic Review, and enabling non-State stakeholders to take a proactive role in accelerating implementation rather than reporting alone. • A defined route by which stakeholder evidence — from civil society, affected communities, technical experts and regulators — formally and equitably enters the Dialogue's deliberations. • A framework for regularly reviewing the implementation of the Dialogue's prioritised set of follow-up workstreams, and any other voluntary or statutory instruments that the Dialogue recommends, and take decisions necessary to promote the effective implementation of the Dialogue, including an agile delivery of this recurring mechanism after the initial sessions till 2027. In short, success should be measured by adoption, institutional readiness and operational capacity, not only by participation or consensus language. Without this architecture, the Dialogue risks becoming a recurring symbolic event whose recommendations are interpreted differently by each Member State and implemented unevenly.

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

We support the four priorities reflected in the joint submission. We further wish to emphasise that AI capacity-building, although not selected within the four-priority limit, is the indispensable enabling condition for all of them. These priorities cannot be implemented without: • Trained regulators, parliamentarians, judges, auditors, procurement officials, incident responders, researchers and civil society and affected communities able to engage meaningfully with AI systems and policies; and • Practical governance infrastructure such as policy observatories, audit pipelines, foresight tools and institutional routines. Because AI systems are often designed in a few Northern jurisdictions but deployed across the world and especially across the Global Majority, a capacity gap in one jurisdiction can quickly become a detection, response, mitigation and remedy gap for others. A regulator unable to assess a frontier model, a procurement office unable to evaluate algorithmic bias, or a judge unable to adjudicate an automated decision creates risks that cross borders and weakens AI governance efforts more broadly. Capacity-building is therefore a shared governance priority for all Member States and a precondition for any meaningful interoperability across national AI governance regimes. We urge the Dialogue to treat capacity-building as cross-cutting in agenda design, in the structure of follow-up workstreams, and in the allocation of resources for Member State and stakeholder participation.

We emphasise one cross-cutting issue that, in our view, deserves explicit elevation: the role of Global Majority expertise, lived experience and local wisdom as foundational sources of governance knowledge, not merely stakeholder perspectives. Today, global AI debates are too often shaped by narrow commercial and national-security framings produced in a small number of well-resourced institutions. Lived realities, locally generated research, and Global Majority expertise remain underrepresented in how risks are defined, prioritised and addressed. This imbalance allows certain systems or community harms - including institutional erosion, community fragmentation, cultural and linguistic erasure, and long-term social harm - to be overlooked because they do not register in the dominant analytic frames. Treating these perspectives as governance knowledge, rather than as input requiring translation, has practical implications: • Risk taxonomies and safety priorities should incorporate harms identified by affected communities and Global Majority researchers, alongside technical risk analyses. • Regional research, harm-mapping and language-specific evaluation work should be commissioned and cited with the same weight as work produced by frontier-model developers. • Outputs of the Dialogue (declarations, workstream reports) should be reviewed for whether they reflect this knowledge, not only whether they have been consulted. This framing is consistent with the Dialogue's universal-forum mandate: governance built on partial knowledge will fail unevenly, and the burden will fall on those whose perspectives were excluded from its design.

The joint submission identifies several harms arising from current governance gaps. We add a regional perspective from Southeast Asia, where high adoption combines with uneven institutional capacity and cross-border economic exposure, making early action especially urgent. Drawing on our work across regional policy tracking, capacity-building, worker-signal research and governance dialogues, we see three linked gaps. 1) Capacity gaps in the Global Majority. Comparable gaps exist in upper-income jurisdictions, but are exacerbated for the Global Majority by economic inequality and asymmetric power. In our region, regulators lack frameworks to evaluate frontier models pre-deployment; incident-response teams cannot triage AI harms at scale; procurement officers cannot assess algorithmic bias; parliaments struggle to draft contextually appropriate AI laws; and judges adjudicating automated decisions lack technical literacy. These gaps are most acute where social stakes are highest — healthcare, justice, financial inclusion, education, social-service delivery, border regimes and public administration. Where frontier-model providers set deployment terms unilaterally, governments without domestic capacity cannot negotiate from an informed baseline, and lawmaking risks being outsourced to foreign consultants or corporate stakeholders. 2) Labour transitions. Displacement falls hardest on labour-intensive sectors central to Southeast Asian economies, including call centres and IT back-office functions, platform work and shared services. From our own research, early warning signs may appear first as workflow redesign, KPI pressure, thinner junior pathways, informal AI use and weaker worker support - not only as immediate displacement. Coordinated transition frameworks — building on existing ILO instruments, with cross-border reskilling commitments and clear augmentation/substitution distinctions reflected in labour regulation, corporate disclosure and procurement — would be more effective than fragmented national responses. 3) AI-mediated social harms and information integrity. AI chatbots facilitate self-harm content for minors, and AI-enabled sextortion targets women (especially female politicians) and girls. Existing frameworks have not adequately evaluated guardrails by non-State actors against targeted disinformation, including climate dis- and misinformation. Deteriorating information environments weaken public trust and institutions, deepening fragmentation in contexts already facing polarisation, conflict or low public trust. 4) Environmental concerns. AI infrastructure's energy and water demand competes directly with Paris-Agreement-aligned renewable-energy transitions and burdens communities near data centres in resource-stressed Global Majority regions, deepening power imbalances with resource-rich industry actors.

We support the joint submission's framing of the Dialogue as a coordination focal point. We add that the Dialogue should explicitly equip itself to enable rapid, coordinated international response when AI-related crises emerge, including risks at the upper end of severity such as catastrophic, systemic or potentially existential-scale harms from advanced AI capabilities, but the practical focus should be operational readiness: verification channels, contact points, escalation paths, tabletop rehearsal and access to independent technical expertise. Relevant crisis types may include: • Misuse of advanced AI capabilities; • AI-assisted cyber incidents; • Systemic infrastructure failures affecting financial systems, public services, energy or shared digital infrastructure; • Public-health risks arising from AI-mediated research, diagnostic or treatment decisions; • Information-manipulation campaigns of cross-border scale; • Unexpected harms when capability gains outpace evaluation and institutional response capacity. Without an established coordination mechanism, Member States respond in isolation, often through national-security or commercial channels that magnify rather than contain fragmentation. We encourage the Dialogue to establish a standing readiness function — proportionate, light-touch and built on existing UN coordination mechanisms — that can mobilise scientific, technical and diplomatic capacity quickly and equitably across all regions. We acknowledge that, because AI cuts across multiple thematic areas and UN institutional jurisdictions, this function may need to be designed as a novel arrangement, drawing for example on the evolving Resident Coordinator system rather than relying solely on existing crisis machinery (UNSC, UNOCHA).

In addition to the international frameworks listed in the joint submission, the Dialogue should connect with regional public-interest infrastructure that demonstrates how implementation can be made tractable in Global Majority contexts. From Southeast Asia, examples include: • SEA Observatory — a live public-interest governance infrastructure stack that combines a source-linked regional AI policy record, citation-first multilingual querying, and scenario-based decision rehearsal so Southeast Asian institutions can track developments, verify materials, and test governance choices before failures occur. • SEA-LION — regional-language large language model work that strengthens linguistic and cultural inclusion in AI development. • RAISE SEA — regional preparedness work using worker-signal research, governance-readiness indicators, transition analysis, and scenario tools to anticipate how AI adoption may reshape labour markets, institutional capacity, social resilience, and economic disruption. • AI Verify — Singapore's open-source testing framework, illustrating practical assurance tooling that other jurisdictions can adapt. • Regional harm-mapping initiatives documenting AI-related harms in context. Investing in similar public-interest infrastructure, in coordination with regional bodies, can reduce duplication, lower verification costs, strengthen linguistic and cultural inclusion, and ensure that governance information is actively used to inform implementation — not left as static documentation. The Dialogue's added value is to recognise such initiatives, identify equivalents in other regions, and curate a shared bench of good practices that can be adopted or adapted across regions.

We support the multistakeholder mechanisms in the joint submission. We add that the Dialogue should invite Member States to designate or identify lightweight national AI governance focal points or coordination contacts, supported with appropriate technical assistance, information-sharing and, where needed, financial resources. These contacts could draw lessons from comparable focal-point practices in climate governance, including under UNFCCC processes and Paris Agreement implementation contexts, while remaining adapted to AI's cross-sectoral nature. Their functions could include: • Identifying areas for international cooperation and opportunities for synergy with other UN processes; and • Routing Dialogue outputs and evidence requests to relevant ministries, regulators, parliaments, standards bodies and public-interest stakeholders; and • Coordinating the preparation of an AI Governance chapter in national communications, with relevant contact information including websites. Focal points are intended to complement, not replace, the sectoral and specialised agencies through which Dialogue outcomes will continue to flow in each country, and to provide a shared coordination point that supports — rather than substitutes for — national readiness.

We support the joint submission's identification of underrepresented voices. We further emphasise the following points that we would have wished to make more prominently in the joint text: • Youth. Young people, who will live longest with the consequences of AI deployment decisions made today, are almost entirely absent from policy-making discussions. A dedicated youth advisory track, connected where appropriate to existing UN youth engagement mechanisms, would enable young people to contribute to agenda-setting, evidence submission and follow-up. • Specific affected groups. Workers whose tasks are being redesigned, women and girls targeted by AI-enabled abuse, children exposed to unsafe AI companions, migrants and gig workers, people with disabilities, gender and sexual minorities, and low-income users hold irreplaceable perspectives that should inform governance. • Resourcing. Meaningful participation requires resources. We urge the Dialogue to support participation by stakeholders without institutional funding, including travel, translation and connectivity support, and to consider time-bound stakeholder rapporteurs or thematic facilitators to synthesise inputs from civil society, affected communities, workers, youth, women and girls, and local-language researchers, in structured exchange with the Co-Chairs and the Independent International Scientific Panel on AI. We further urge the Dialogue to invite Member States to designate financial resources for national coordination with non-State stakeholders, as appropriate.

In addition to the engagement formats reflected in the joint submission, we propose that the Dialogue launch a Global Commitment Incubator and Implementation Accelerator: a voluntary, multistakeholder format led by the Co-Chairs to help Member States and stakeholders translate Dialogue outcomes into: • Concrete governance commitments; • Implementation roadmaps; • Capacity-needs assessments; and • Annual progress updates. The Incubator format would be especially valuable for States and stakeholders that endorse Dialogue priorities but lack the technical, legal or institutional infrastructure to operationalise them on national timelines. It would convert declaratory progress into time-bound, measurable implementable progress, while preserving the voluntary, non-prescriptive character of the Dialogue. AI Safety Asia could contribute Southeast Asian use cases, policy-tracking inputs, decision-rehearsal formats, capacity-building designs and transition-evidence methods to such an accelerator, while working with Member States, regional bodies, civil society, researchers and technical partners. This would help ensure that Dialogue outcomes are not only discussed globally, but tested, adapted and implemented in real institutional settings.

Government approaches reflect diverse policy priorities and institutional contexts. For example: Brazil: rights-based approach through PL 2338/2023 (pending enactment); prohibits excessive- risk AI; establishes rights to explainability and human review. China: comprehensive regulatory approach through PIPL (2021), Algorithm Recommendation Provisions (2022), and Interim Measures on Generative AI (2023); mandates algorithmic transparency, content moderation and security assessment; actively advancing international governance through Global AI Governance Initiative. EU: legally binding risk-based governance through EU AI Act; prohibits "unacceptable risk" practices including manipulation causing harm, public-sector social scoring and certain biometric identification; mandates conformity assessments, post-market monitoring and human oversight. India: "techno-legal" approach combining baseline data protection safeguards, sectoral regulation and technical controls; emphasizes inclusion and digital public infrastructure through IndiaAI Mission rather than single overarching AI law. Singapore: innovation-enabling approach through Model AI Governance Framework (including 2024 Generative AI guidance) and open-source AI Verify testing framework; emphasizes voluntary adoption and interoperability. South Africa: human-centric, ethics-first approach through National AI Policy Framework (2024), emphasizing Ubuntu principles, inclusion and alignment with AU Continental AI Strategy. South Korea: innovation-led approach through AI Basic Act (2026) with risk-based obligations for "high-impact AI", active international coordination through AI Safety Institute and Seoul Summit. UAE: innovation-first positioning through the National AI Strategy 2031 and dedicated AI Minister (since 2017); substantial investment in domestic capacity and active international AI engagement. USA: innovation-first approach without federal omnibus AI legislation; combines voluntary NIST AI Risk Management Framework, targeted federal statutes (TAKE IT DOWN Act) and active state legislation (California SB 53, Texas Responsible AI Governance Act, New York RAISE Act, Virginia SB 384 / HB 797). Using key building blocks of AI governance as a comparative framework, the Dialogue can identify which are working in practice, where capacity gaps persist and which safeguards should become interoperable across jurisdictions: https://global-governance.ai/. The Dialogue's added value is to compare which building blocks are working in practice, where capacity gaps persist, and which safeguards should become interoperable across jurisdictions.