UNIDO

• In my opinion, a successful first Global Dialogue on AI Governance should achieve several key outcomes: (1) The dialogue must meaningfully include voices from the Global South, civil society, academia, and smaller nations. Success means no major stakeholder group feels sidelined, and developing countries see their concerns (capacity gaps, digital divides, data sovereignty) reflected in the outcomes
• (2) Participants don't need to agree on solutions yet, but they should leave with a common understanding of the core challenges: safety risks, concentration of power, cross-border enforcement gaps, and the tension between innovation and precaution. A shared vocabulary and taxonomy of risks would be a concrete, valuable deliverable
• (3) Rather than another aspirational declaration, success means establishing practical next steps — whether that's an interoperability framework for national AI regulations, mutual recognition agreements for safety evaluations, or a standing technical body for information sharing on frontier model risks. Even modest institutional commitments would mark real progress
• (4) Bridging the governance fragmentation. Currently, AI governance efforts are scattered across the EU AI Act, the Hiroshima Process, the Bletchley Declaration, the UN Advisory Body, and various bilateral deals. A successful dialogue would map these initiatives against each other and identify where they complement or conflict, reducing redundancy and confusion
• (5) Perhaps most importantly, the dialogue should establish itself as a credible, ongoing forum, rather than a one-off event. If it earns enough buy-in that major actors (including leading AI companies) commit to participate in future rounds and to transparency measures, that alone represents a meaningful shift toward accountable global governance. In summary, the bar for "success" should be realistic: durable process over premature consensus.

• AI capacity-building
• Open-source software, open data and open AI models
• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI

My selection reflects priorities that I believe are both foundational and actionable for advancing equitable AI governance: (1) Safe, secure and trustworthy AI is the prerequisite for all other governance objectives. Without robust safety frameworks, covering frontier model evaluation, adversarial robustness, and reliability assurance, downstream efforts in rights protection or capacity-building rest on unstable ground. As a researcher focused on intelligent computing system performance evaluation, I see firsthand that rigorous, reproducible benchmarking methodologies are essential to substantiate trustworthiness claims and move beyond self-reported safety assurances. (2) AI capacity-building is urgent because the current global AI landscape is deeply asymmetric. A handful of nations and corporations concentrate the computational infrastructure, talent, and data resources needed to develop and deploy advanced AI systems. Without deliberate investment in capacity-building, including access to computing resources, training programs, and institutional knowledge transfer, the governance dialogue risks becoming a conversation among those who already hold power, about technologies that others can only passively receive. Our work within the UNIDO ecosystem reinforces this conviction: developing countries need not just access to AI tools, but the technical capacity to evaluate, adapt, and govern them on their own terms. (3) Social, economic, ethical, cultural, linguistic and technical implications of AI must be addressed holistically. AI systems do not operate in a vacuum; their deployment reshapes labor markets, cultural production, and linguistic diversity. Governance frameworks that treat AI as a purely technical matter will inevitably fail to anticipate or mitigate these broader societal impacts. (7) Open-source software, open data and open AI models represent a critical lever for democratizing AI development and reducing dependency on proprietary ecosystems. Open approaches lower barriers to entry, enable independent auditing and reproducibility, and foster innovation across diverse contexts. They are also a practical mechanism for advancing the capacity-building and trustworthiness goals outlined above.

Here are several cross-cutting issues that deserve explicit attention beyond the listed themes. (1) AI infrastructure concentration and compute governance. The listed themes address software, data, and models, but largely overlook the material foundation of AI: computational infrastructure. Access to large-scale GPU clusters, energy supply, and semiconductor supply chains fundamentally determines who can build, train, and deploy frontier AI systems. Governance discussions that ignore compute concentration risk addressing symptoms while leaving the structural power asymmetry untouched. International frameworks should consider equitable access to high-performance computing resources as a governance priority in its own right; (2) Evaluation and benchmarking standardization. There is currently no internationally recognized methodology for assessing AI system capabilities, safety margins, or performance claims. Vendors self-report benchmarks under inconsistent conditions, making meaningful comparison, and therefore meaningful regulation, nearly impossible. A cross-cutting effort to establish shared evaluation protocols, reproducible testing environments, and independent auditing standards would strengthen every thematic area listed, from trustworthiness to interoperability of governance approaches; (3) Environmental sustainability of AI development. The energy and water consumption of training and operating large-scale AI models is growing rapidly, yet environmental impact remains largely absent from governance frameworks. This issue cuts across economic, ethical, and technical dimensions and warrants dedicated attention, particularly as developing countries are asked to build AI capacity without exacerbating climate vulnerabilities; (4) AI governance for scientific research. AI is increasingly embedded in scientific discovery such as from drug design to climate modeling. The governance implications for research integrity, reproducibility, and equitable access to AI-augmented research tools represent a distinct challenge that does not fit neatly into any single listed theme but has profound long-term consequences for global knowledge production. Above issues are not peripheral; they shape the conditions under which all listed thematic areas will succeed or fail.

As a researcher affiliated with Tsinghua University's Department of Computer Science and technical specialists within UNIDO, I observe governance gaps affecting both China's AI ecosystem and the broader international development sector. 1. Challenges. Fragmented evaluation standards pose a direct obstacle to our work. China has developed its own AI safety and evaluation frameworks, while the EU, the US, and other jurisdictions pursue divergent approaches. This fragmentation creates compliance burdens for cross-border research collaboration and makes it difficult to establish mutual recognition of safety assessments. For our work on intelligent computing system performance evaluation, the absence of internationally harmonized benchmarking protocols means that capability and safety claims remain difficult to verify and compare across systems and jurisdictions. Within the UNIDO context, we witness a stark capacity gap among developing member states. Many countries that UNIDO serves lack the computational infrastructure, technical talent, and institutional frameworks to meaningfully participate in AI governance, let alone develop or adapt AI systems to local needs. Governance discussions risk producing standards that these countries can neither implement nor influence, deepening existing inequalities. The tension between open-source AI promotion and security concerns also presents a significant challenge. China is home to a vibrant open-source AI ecosystem, yet evolving export controls and geopolitical friction increasingly constrain the free flow of models, data, and computing resources across borders, undermining the collaborative potential that open approaches offer. 2. Opportunities. China's substantial investment in AI research and infrastructure, combined with UNIDO's mandate to support inclusive industrial development, positions us to bridge the gap between frontier AI development and practical deployment in developing economies. We see an opportunity to contribute evaluation methodologies, capacity-building frameworks, and open technical resources that can serve as shared public goods, provided governance structures create the enabling conditions for such collaboration rather than further fragmentation.

• The AI Dialogue can serve as a uniquely legitimate platform for advancing international cooperation, provided it fulfills several critical functions: (1) Establishing a common technical language. Current AI governance discussions suffer from terminological confusion, terms like "frontier models," "high-risk systems," and "trustworthy AI" carry different meanings across jurisdictions. The Dialogue can convene technical and policy experts to develop shared definitions and taxonomies, which is a precondition for any substantive regulatory coordination. Without this foundational alignment, interoperability of governance approaches remains aspirational
• (2) Creating a neutral space for bridging geopolitical divides. AI governance is increasingly entangled with strategic competition, particularly between major powers. The UN-anchored Dialogue offers a multilateral setting where technical cooperation can proceed even when bilateral political channels are strained. This is especially valuable for sustaining collaboration on safety research, evaluation standards, and incident reporting — areas where collective action benefits all parties regardless of competitive dynamics
• (3) Amplifying underrepresented voices in norm-setting. Many countries and sectors affected by AI deployment have had minimal input into the governance frameworks being established by early movers. The Dialogue can institutionalize mechanisms, dedicated capacity-building sessions, technical assistance programs, structured consultation processes, that ensure developing countries and specialized UN agencies like UNIDO are not merely observers but active contributors to norm formation
• (4) Coordinating the existing governance landscape. The proliferation of initiatives, the Hiroshima Process, the EU AI Act, the Bletchley Declaration, bilateral agreements, risks duplication and inconsistency. The Dialogue can serve as a mapping and coordination mechanism, identifying complementarities and gaps across existing frameworks without attempting to replace them
• (5) Building iterative trust. Perhaps most importantly, the Dialogue should prioritize process over premature consensus. Regular convenings, transparent reporting, and incremental commitments will build the institutional trust necessary for deeper cooperation over time.

The AI Dialogue should not operate in isolation but rather serve as a connective layer across the already substantial, yet fragmented, landscape of AI governance initiatives. 1. Existing initiatives to build upon. The UN Secretary-General's High-Level Advisory Body on AI has laid important groundwork in identifying governance gaps and proposing institutional options. The Dialogue should treat its recommendations as a starting point rather than reopening foundational debates. Similarly, the Global Digital Compact provides a broader normative framework within which AI-specific governance commitments can be anchored and operationalized. At the plurilateral level, the Hiroshima AI Process and the Bletchley AI Safety Summit series have advanced discussions on frontier model safety among like-minded nations. However, their limited membership constrains legitimacy. The Dialogue can extend these conversations to the full UN membership, ensuring that safety norms reflect genuinely global input. Regional regulatory frameworks, notably the EU AI Act, China's evolving AI regulations, and emerging approaches across Africa, ASEAN, and Latin America, represent diverse governance philosophies. The Dialogue should systematically map these approaches to identify areas of convergence suitable for mutual recognition or interoperability agreements. Within the UN system, specialized agencies bring critical sectoral expertise. UNIDO contributes perspectives on AI for inclusive industrial development and capacity-building in developing economies. UNESCO's Recommendation on the Ethics of AI provides an ethical reference framework. The ITU's AI for Good platform and standardization work offer technical coordination infrastructure. 2. Added value of the Dialogue. The Dialogue's unique contribution lies in three areas: universal legitimacy that no smaller grouping can claim; the convening power to connect technical, political, and development communities that rarely interact; and the institutional continuity to sustain cooperation beyond individual presidencies or political cycles. Its role is not to duplicate existing mechanisms but to make them collectively coherent and globally accountable.

A successful AI Dialogue requires structured participation from diverse stakeholders, each contributing distinct expertise and legitimacy. Member States should bring concrete national experiences, both regulatory experiments and implementation challenges, rather than reiterating general principles. Sharing what has worked and what has failed in domestic AI governance provides the empirical foundation for meaningful international coordination. Besides, the technical and academic community, including institutions like Tsinghua University, can contribute rigorous evaluation methodologies, independent safety assessments, and evidence-based analysis that ground policy discussions in technical reality. Researchers are uniquely positioned to identify gaps between governance aspirations and actual system capabilities. What's more, UN specialized agencies such as UNIDO, UNESCO, and the ITU should contribute sectoral expertise and channel the needs of their constituencies — particularly developing countries, into the Dialogue. Their operational presence in member states provides an implementation pathway that standalone governance forums lack. Also, the private sector, including both frontier AI developers and downstream deployers, should contribute transparency commitments, participate in evaluation exercises, and share operational insights on risk management. Their engagement must be substantive, not performative. On the other hand, the Dialogue should adopt a multi-track structure: a high-level political track for norm-setting and commitment-building, complemented by technical working groups organized around specific thematic priorities. These working groups should produce concrete deliverables, shared evaluation protocols, capacity-building toolkits, interoperability mapping reports, between sessions. We recommend establishing an open intersessional consultation mechanism that allows stakeholders unable to attend in person to contribute written inputs and technical evidence. Regular progress reviews against stated objectives would ensure accountability. The Dialogue should operate on an annual cycle with published agendas and outcome documents, building institutional memory and enabling iterative advancement rather than repetitive declaration.

Current global AI governance discussions exhibit significant representation gaps that risk producing frameworks disconnected from the communities most affected by AI deployment. Developing countries and least developed countries remain structurally marginalized. While they participate formally in UN processes, many lack the dedicated AI policy units, technical expertise, and travel budgets necessary for substantive engagement. Their participation often remains reactive rather than agenda-setting, meaning governance norms are shaped primarily by the small number of nations that both develop and regulate frontier AI systems. The technical research community from the Global South is particularly absent. AI governance discussions are disproportionately informed by researchers affiliated with institutions in North America, Europe, and a handful of East Asian countries. Researchers from Africa, Southeast Asia, Central Asia, and Latin America, who understand local deployment contexts, infrastructure constraints, and societal impacts, are rarely consulted in norm-setting processes. What's more, workers and labor communities displaced or transformed by AI adoption have minimal structured representation. Their perspectives on automation, workplace surveillance, and algorithmic management are essential to the social and economic dimensions of governance yet are typically mediated through secondary advocacy rather than direct participation. Finally, linguistic and cultural minorities face compounding exclusion. AI systems predominantly serve major languages, and governance discussions conducted primarily in English further marginalize communities whose languages, cultural practices, and knowledge systems are underrepresented in both training data and policy forums. The Dialogue should establish funded fellowship programs enabling delegates from underrepresented countries and communities to participate meaningfully. Regional preparatory consultations conducted in local languages can surface priorities before global sessions. A dedicated technical assistance facility, potentially operated through agencies like UNIDO, could help developing countries build permanent AI governance capacity rather than relying on episodic participation. Remote participation infrastructure with professional interpretation should be standard, not optional.

Traditional panel-and-plenary formats have proven insufficient for generating actionable outcomes in multilateral technology governance. The AI Dialogue should experiment with formats designed to produce concrete results rather than parallel monologues.(1) Structured technical demonstrations. Rather than debating AI capabilities in the abstract, sessions should incorporate live demonstrations of AI systems — including their failure modes. Showing policymakers how evaluation methodologies work in practice, how models behave under adversarial conditions, and how safety benchmarks are conducted would ground governance discussions in observable reality. This bridges the persistent gap between technical and policy communities.(2) Problem-centered working sprints. Instead of thematic panels, organize time-boxed collaborative sessions where mixed teams of policymakers, researchers, developers, and civil society representatives work together on specific governance challenges — such as drafting interoperability criteria for national AI registries or designing a shared incident-reporting template. Producing tangible draft outputs within the Dialogue itself creates momentum and accountability. (3) Reverse consultation format. Invite developing country delegations to present their specific governance needs and deployment contexts first, then ask frontier AI developers and advanced-economy regulators to respond with concrete proposals. This inverts the typical dynamic where norms are proposed by powerful actors and passively received by others. (4)Scenario-based stress testing. Facilitate structured exercises where participants collectively work through realistic governance scenarios — a cross-border AI safety incident, a capability breakthrough requiring rapid regulatory response, or a dispute over algorithmic bias in an international development program. These exercises reveal governance gaps more effectively than abstract discussion.

Several existing initiatives offer instructive models for effective AI governance, each addressing different dimensions of the challenge.(1) Evaluation and standardization frameworks. China's national AI standardization efforts, coordinated through TC260 and complemented by sector-specific regulations on algorithmic recommendation, deepfakes, and generative AI, demonstrate an iterative regulatory approach that addresses specific applications rather than attempting comprehensive legislation prematurely. This targeted methodology allows governance to evolve alongside rapidly advancing technology. Similarly, NIST's AI Risk Management Framework in the United States provides a voluntary, structured approach to identifying and mitigating AI risks that has gained traction across sectors precisely because of its flexibility. (2) Regulatory interoperability experiments. The EU AI Act represents the most ambitious comprehensive regulatory framework to date. While its extraterritorial reach raises legitimate concerns, its risk-tiered classification system offers a replicable structural model. The emerging dialogue between EU and international counterparts on mutual recognition of conformity assessments provides a practical template for governance interoperability. (3) Open-source ecosystem governance. Initiatives such as Hugging Face's model card framework and the BigScience project's RAIL licensing approach demonstrate that community-driven governance mechanisms can embed responsible use norms directly into model distribution pipelines without requiring top-down regulation. China's thriving open-source AI ecosystem, including models from institutions like Tsinghua University, shows that openness and safety governance can coexist productively.