Beijing Institute of AI Safety and Governance

The success of the inaugural Dialogue should not be measured by the number of statements made, but by whether it can build a practical bridge between principles and implementation. A successful first Dialogue should deliver at least four outcomes. First, it should identify a clear, limited and actionable set of international cooperation priorities, especially in AI safety and security testing, governance capacity-building, interoperability of governance approaches, and human-rights-based accountability. Second, it should set out a follow-up work arrangement from Geneva in 2026 to New York in 2027, so that the Dialogue becomes a continuing process rather than a one-off event. Third, it should provide genuinely usable outputs for countries and institutions with limited resources, such as risk assessment templates, evaluation guidance and capacity-building support. Fourth, it should create a trusted space where governments, researchers, companies and civil society can share not only best practices, but also candidly discuss questions that remain unclear, unresolved or not yet effective.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches

These four priorities are important because the central tension in AI governance today lies between rapidly expanding technical capabilities and relatively limited institutional readiness. First, safe, secure and trustworthy AI is foundational. Current systems may appear aligned in many settings, but this does not mean that value alignment has been solved; in complex, adversarial or high-risk contexts, systems may still deviate, display deceptive or superficial compliance, or behave unpredictably. Second, capacity-building is urgent because the AI divide is not only a gap in compute, data and talent, but also a gap in testing capacity, risk assessment capacity and governance implementation capacity. Third, social, economic, ethical, cultural, linguistic and technical impacts must be considered together, because AI's broad effects differ across social contexts and cannot be addressed through a single technical lens. Finally, links among risk management frameworks in different countries and institutions should be considered in order to support multi-stakeholder collaborative governance.

Yes. Several cross-cutting issues deserve explicit attention. First, one prominent risk is the public's misperception of the boundaries of AI capability. More people are beginning to project consciousness, intention or even moral agency onto current systems; this over-anthropomorphization is inconsistent with the current scientific understanding. If not addressed, it may lead to over-trust, excessive disclosure of sensitive information, and inappropriate delegation of judgment that should remain with humans. Second, AI is being introduced into governance workflows themselves, including ethics risk screening, document triage and compliance review. This may improve efficiency, but it can also create new challenges such as automation bias, hidden errors and the templating of expert judgment. It remains unclear what enabling benefits AI ethics experts may gain from AI, and what new challenges they may face; this requires further exploration.

The most immediate impact is that deployment is moving faster than governance readiness. In many sectors, large models and related systems are already entering pilot use, procurement and routine workflows, while supporting testing standards, documentation requirements, ethics review procedures and responsibility allocation are still taking shape. This leaves researchers, policymakers and public institutions without enough comparable evidence when assessing risks. At the same time, there are encouraging developments. Relevant ministries and specialized institutions are advancing national-level AI safety and security work. Under top-level institutional arrangements such as national science and technology ethics committees, and guided by relevant science and technology ethics review rules, many institutions and companies are actively establishing ethics committees, with practices gradually extending toward municipal, provincial and national ministry levels. Some practices are also beginning to use AI to assist ethical risk screening for language models. These explorations are meaningful, but their effectiveness, replicability and limitations still require continuous evaluation.

The Dialogue's most important role is to help translate the international community's principled consensus on AI governance into a sustained, comparable and collaborative process of practice. It can help stakeholders identify areas where minimum common ground is already possible, and help countries better understand differences in cultural contexts, institutional traditions and governance resources. It can also enable countries to share cross-cultural governance experience, tools and platforms with one another, thereby strengthening global AI governance preparedness, especially by supporting low- and middle-income countries in improving their governance systems. In addition, the Dialogue can promote shared scientific understanding of frontier AI risks, risk prevention and collaborative governance.

• Future work should strengthen links with existing international initiatives rather than starting from scratch. Important foundations include UNESCO's Recommendation on the Ethics of Artificial Intelligence and its implementation work
• follow-up arrangements within the UN system on AI governance
• ITU and AI for Good technical cooperation platforms
• cross-cultural cooperation and governance platforms such as the World Internet Conference (WIC. https://www.wicinternet.org/), the World Artificial Intelligence Conference (WAIC. https://www.worldaic.com.cn/), the International Research Network on AI Development and Governance (AIR Net. https://air-net.ai/), and the Alliance of National and International Science Organizations for the Belt and Road Regions (ANSO. https://www.anso.org.cn/)
• policy exchange mechanisms such as the OECD and GPAI
• and standards work by ISO, ACM and IEEE. National practices in safety testing, regulatory pilots, deployment of AI in public-sector services and ethics committee development should also be included in comparative learning. The Dialogue's added value is to turn work dispersed across different institutional spaces into shared knowledge that is more interoperable, understandable and accessible. At the UN level, a mechanism is needed to coordinate implementation, curate reusable tools, identify capacity gaps, promote cross-cultural resource-sharing, and help countries find common ground for coordinated governance.

The AI Dialogue should uphold multi-stakeholder governance and agile governance, and clarify the role positioning, responsibility boundaries and degrees of responsibility of different actors at the institutional level. States should participate in multilateral consultation, compliance review and regular responsibility reporting. Enterprises should fulfill transparency obligations and undergo algorithmic safety audits and data protection compliance assessments. Research institutions and international organizations should rely on independent ethics review and technical assessment to strengthen the normative and compliant conduct of research and governance processes. Structurally, a multi-level coordination system can be established with global common rules as the foundation, regional norms as support and industry standards as a supplement, promoting mutual recognition and linkage between international norms and domestic legislation, as well as between general requirements and rules for specific scenarios. At the same time, implementation should be strengthened through a combination of soft and hard mechanisms: ethical principles and industry codes can form voluntary commitments, while international treaties, regional agreements and regulatory cooperation memoranda can reinforce compliance guarantees. Dynamic adjustment mechanisms should also be used to respond to rapid technological change

Voices from low- and middle-income countries and regions with relatively limited governance capacity are often most affected by external technologies and platform rules, yet lack the resources to shape the international agenda. Frontline users and institutions, including local governments, schools, hospitals and labor regulators, have direct knowledge of deployment risks and implementation difficulties, but are often not sufficiently included. Children, youth, persons with disabilities and other affected groups are often consulted only after rules have largely taken shape. Addressing this requires more than expanding invitation lists. It also requires financial support, regional preparatory consultations, hybrid participation mechanisms, and institutional arrangements that allow local experience to enter agenda-setting and outcome documents. The Artificial Intelligence for Children: Beijing Principles(https://ai-ethics-and-governance.institute/artificial-intelligence-for-children-beijing-principles/), for example, calls on all sectors of society to attach great importance to AI's impact on children.

Meaningful and active participation is best supported by regular, multi-stakeholder mechanisms for joint research and capacity-building around frontier AI issues. The World Internet Conference (WIC) Specialized Committee on Artificial Intelligence provides a useful reference: it establishes a thematic platform involving international organizations, think tanks, research institutes, professional associations and industry, and carries out joint research, thematic seminars, outcome sharing and initiative releases in areas such as AI safety and governance, standards development and industrial application. Such mechanisms can turn one-off conference statements into sustained cooperation, and turn dispersed views into shared research outputs, policy tools and practical cases. For the UN AI Dialogue, regional joint research groups and capacity-building projects could further support low- and middle-income countries in participating in risk assessment, ethics review, standards translation and governance tool development, improving the substance, fairness and sustainability of participation.

Several complementary pathways can be referenced. At the institutional level, China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services (https://www.cac.gov.cn/2026-04/10/c_1777558395078289.htm) address emerging risks such as emotional companionship and virtual intimate relationships, exploring classified and graded governance, life-cycle governance and protection of minors. The Measures for Ethical Review and Services for AI Science and Technology (Trial) (https://www.miit.gov.cn/zwgk/zcwj/wjfb/tz/art/2026/art_c5039010f5d24e1593152a9355f9c51c.html) embed ethics requirements throughout scientific research and technology development. The AI Safety Governance Framework 2.0 (https://www.cac.gov.cn/2025-09/15/c_1759653448369123.htm) provides tools for risk classification, risk grading and dynamic governance. At the platform level, the AGILE Index, and the Global AI Governance Online Observatory (https://agile-index.ai/)can compare countries' governance preparedness and support evidence-based decision-making. At the mechanism level, the WIC Specialized Committee on Artificial Intelligence (https://cn.wicinternet.org/node_164157.htm) promotes regular international cooperation through its Standards Program, AI Safety and Governance Program, and Industry Program. At the technical level, ForesightSafety Bench https://foresightsafety-bench.beijing-aisi.ac.cn/)can provide operational tools for model safety evaluation, risk identification and governance capacity-building. Together, these practices demonstrate a governance pathway that combines rules, assessment, platforms and technical benchmarks.