Skip to content

World Internet Conference AI Committee Report: Advancing a Global AI Safety and Governance Framework for the Benefit of Humanity

Advancing a Global Framework for AI Safety and Governance for the Well-being of Humanity

arXiv: Alignment, Deception & Oversight· arXiv:2611.13968· Zeng Yi, Seán Ó hÉigeartaigh, Wang Zhengqi, Lu Enmeng, Chen Yu, Cao Gongce, Guo Xiaoyang, Kang Yanrong, Han Kaiyu, Fan Jinyu, Xie Jiawei, Han Zhengqiang, Wang Jin, Huangfu Cunqing, Bao Aorigele, Dame Wendy Hall, Lin Lin, Duan Weiwen, Wang Rong, Zhang Junlin, Tang Xinhua, Vincent C. Müller, Chen Jingjing, Li Na, Cheng Kai, Sebastian Sunday Grève, Danil Kerimi, Bernd Holznagel, Anna Abramova, Jimena Sofia Viveros Alvarez, Edson Prestes, Shen Juncheng, Liu Yongmou, Zhang Jiji, Nada Laabidi, Yao Xin, Zhou Kai, George Chen, Fu Chunhui, Liu Xiaochun, Guo Sumin, Hu Naying, Qiao Qian, Meng Wei, Zhang Linghan, Tan Zhixing, Helen Meng, Qiao Yu, Bu Yuyan, Fan Wei, Wang Xingguang, Wang Feng, Gong Xinqi, Xia Wenhui, Wang Mengyin, Tao Feng, Hu Yongqi, Charuka Senal Damunupola, Cheng Ming, Liu Weichen, Nirosha Ananda, Peng Tao, Tao Tao, Wang Fengqiong, Wang Jianbing, Wang Tong, Wang Wei, Wang Xin, Wu Shuyan, Yang Yaodong, Yang Zhongliang, John Yeoh, Zhang Rong, Zhou Yuan, Zhang Xueli, Liang HaoChinese Academy of Sciences; Beijing Institute of AI Safety and Governance (Beijing-AISI); University of CambridgeIndexed by this site · 6 days ago· Originally publishedGovernance

Quick read

Other

Based on the paper PDF (AI-generated); refer to the original

A World Internet Conference report reviews the current state of global AI safety governance and proposes a global AI safety governance framework and mechanism recommendations centered on the United Nations.

Questions
AI risks spread across borders while the global governance system is fragmented; differences among countries in development level, governance capacity, and priorities hinder international cooperation, and there is an urgent need to establish a global AI safety governance framework with broad consensus.
Methods
It systematically reviews the governance practices of the United Nations, countries, regions, and international organizations, draws on multilateral governance experience from five domains—nuclear safety, climate, trade, public health, and aviation and maritime—and proposes framework recommendations centered on the United Nations around mechanism design.
Experiments and results
Data cited in the report show that from 2019 to 2024, global AI risk incidents increased from about 400 to more than 7,900; 118 of the 193 UN member states did not participate in the seven major AI governance initiatives; only 6 countries signed all five key international AI safety declarations.
Limitations and future work
The report has no limitations discussion section; its content is a qualitative review of policy documents and governance practices and recommendations, and it contains no experiments, model evaluations, or quantitative validation.
AI summaryThe Artificial Intelligence Committee of the World Internet Conference released a report proposing institutional measures and policy recommendations for establishing a global AI safety and governance framework centered on the United Nations. The report systematically reviews global exploration and practice in AI safety and governance, identifies key issues that need to be addressed, and draws on multilateral governance experience from other domains. Centered on mechanism design, the report explores building a global governance system that ensures safety, promotes inclusiveness, clarifies responsibility and accountability, strengthens coordination, and is authoritative and efficient.

The Artificial Intelligence Committee of the World Internet Conference released a report proposing institutional measures and policy recommendations for establishing a global AI safety and governance framework centered on the United Nations. The report systematically reviews global exploration and practice in AI safety and governance, identifies key issues that need to be addressed, and draws on multilateral governance experience from other domains. Centered on mechanism design, the report explores building a global governance system that ensures safety, promotes inclusiveness, clarifies responsibility and accountability, strengthens coordination, and is authoritative and efficient.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    AI risks spread across borders while global governance is fragmented; the report explores how to build a UN-centered global AI safety governance framework.

    Core question: AI risks have cross-border, cross-domain systemic characteristics, while the global AI safety governance system is fragmented; how to build an actionable global AI safety governance framework based on broad consensus.

    • Problem scenario: The report states that AI risk incidents continue to grow; from 2019 to 2024, globally recorded AI risk incidents increased from about 400 to more than 7,900 (Figure 1), among which incidents in the robustness and digital security, human rights and privacy, and transparency and accountability categories accounted for more than 60%.
    • Shortcomings of existing governance: The report holds that current global governance has four structural limitations—technological breakthroughs are hard to predict, leading to governance lag; countries differ greatly in development level and governance models; stakeholders are diverse while accountability mechanisms are unclear; and factors such as geopolitics weaken collective response capacity (Section 01(c)).
    • Core observation: AI risks inherently transcend national borders, regulatory measures of a single country or region can hardly constrain them effectively, and measures such as export controls and technology blockades by some countries further compress the space for international cooperation (Section 01(a)).
    • Paper output: Drawing on multilateral governance experience from five domains—nuclear safety, climate change, international trade, public health, and aviation and maritime (Table 1)—the report proposes, around the four pillars of ensuring security, ensuring inclusion, clarifying responsibility, and practicing multilateralism, institutional measures and policy recommendations for a global AI safety governance framework centered on the United Nations (appendix).
  2. What related research exists?

    The report reviews AI governance practices of the United Nations system, countries, regional organizations, international organizations, international summits, and academia and industry, and draws on multilateral governance mechanisms from five domains.

    Section 01(b) of the report systematically reviews governance explorations by various global stakeholders, and Section 01(d) analyzes multilateral governance frameworks in other domains.

    UN system

    • The UN Secretary-General's High-level Advisory Body released the interim and final reports of "Governing AI for Humanity"; the UN General Assembly adopted multiple resolutions and decided to establish an independent international scientific panel and a global dialogue mechanism on AI governance.
    • UNESCO released the "Recommendation on the Ethics of Artificial Intelligence" and accompanying assessment tools; ITU held the AI for Good Global Summit; UNICEF released policy guidance on AI for children.

    National and regional levels

    • The U.S. NIST released the "AI Risk Management Framework" and promoted the International Network of AI Safety Institutes; Singapore released two editions of its Model AI Governance Framework; China released the "Global AI Governance Initiative," the "AI Safety Governance Framework (1.0, 2.0)," and the "Global AI Governance Action Plan," and proposed establishing a World AI Cooperation Organization.
    • The EU adopted the "Artificial Intelligence Act," which the report describes as the world's first comprehensive AI legal framework; the Council of Europe introduced the first legally binding international treaty on AI; the Arab League, African Union, ASEAN, and Latin America each released regional AI strategies or declarations.

    International organizations and summit mechanisms

    • The G20, G7 (Hiroshima AI Process), OECD (AI Principles and GPAI), APEC, BRICS, SCO, World Internet Conference, and others released relevant declarations or documents.
    • The series of AI safety summits since 2023 produced voluntary outcomes such as the "Bletchley Declaration," the "Seoul Ministerial Statement," and the "Frontier AI Safety Commitments"; the REAIM Summit focused on responsible AI in the military domain.

    Industry-academia self-regulation and frontier risk frameworks

    • Academia promoted risk consensus through IDAIS and the CAIS "Statement on AI Risk"; at the corporate level, frameworks include Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, Google DeepMind's Frontier Safety Framework, Meta's Outcome-Focused Frontier AI Framework, and the China Artificial Intelligence Industry Development Alliance's "AI Safety Commitment."

    Multilateral governance lessons from other domains (Table 1)

    • In five domains—nuclear safety (IAEA combining verification and technical assistance), climate change (IPCC scientific assessment driving policy), international trade (WTO dispute settlement mechanism), public health (WHO International Health Regulations), and aviation and maritime (ICAO/IMO standards coordination)—the report considers that they offer lessons for AI governance in compliance verification, scientific assessment, dispute settlement, rapid response, and mutual recognition of standards, but the generative capability, autonomy, and evolvability of AI mean existing institutional models cannot simply be copied.

    Report positions itself as: on the basis of integrating the above scattered mechanisms, proposing a global framework design centered on the United Nations that balances security and development inclusiveness.

  3. How does the paper address this problem?

    The report organizes its analysis around the four pillars of "security—inclusion—responsibility—multilateralism," draws on governance experience from five domains, proposes mechanism design recommendations, and summarizes them in an appendix table.

    This is a policy report rather than a technical paper; its "method" is systematic review plus mechanism design: first diagnose risks and governance gaps, then draw on multilateral governance experience from other domains, and finally propose mechanism recommendations organized around four pillars, with an appendix summarizing them in table form.

    Problem diagnosis and framework structure

    • The report summarizes the key issues in global AI governance into four interrelated pillars: ensuring security (primary objective), ensuring inclusiveness (cornerstone of a fair governance structure), clarifying responsibility (support for multi-stakeholder coordination), and practicing multilateralism (value orientation) (Section 01(c)).
    • Each pillar is developed according to the structure "risk/challenge analysis → mechanism design recommendations" (Chapters 02–05).

    Borrowed mechanism design (Section 01(d), Table 1)

    • From the IAEA, borrow a compliance mechanism combining verification and technical assistance; from the IPCC, borrow a dynamic governance model in which scientific assessment drives policy; from the WTO, borrow binding dispute settlement and arbitration; from the WHO International Health Regulations, borrow cross-border information sharing and rapid response; from ICAO/IMO, borrow standards coordination and continuous oversight.
    • The report also notes that AI's generative capability, autonomy, evolvability, and cross-domain penetration make its risk spillovers broader and impact chains longer, so existing institutional models cannot simply be copied.

    Mechanism recommendations for each pillar

    • Ensuring security (Chapter 02): establish a transnational technical monitoring network and risk early-warning collaboration mechanism, a mechanism for dynamic updating and mutual recognition of governance rules, and a co-built ecosystem of safety assessment tools and platforms; promote an internationally agreed risk classification and grading framework, cross-border content governance collaboration (digital watermarking, content provenance), and cross-border law enforcement cooperation; for major global risks, establish a frontier model monitoring and assessment mechanism, an international governance system for critical resources (compute, data, open-source algorithms), and an emergency intervention and joint response mechanism.
    • Ensuring inclusiveness (Chapter 03): establish an open and inclusive multilateral consultation platform, technology sharing and transfer mechanisms, innovative financing models (AI development fund, PPP financing pool), and long-term capacity-building cooperation mechanisms; establish a dynamic compliance assessment and reputation linkage mechanism, a decision-making hearing mechanism, and a differentiated responsibility-sharing mechanism based on development level.
    • Clarifying responsibility (Chapter 04): define respectively the roles and responsibility boundaries of national governments, international organizations, enterprises, research institutions, the public, and civil society; establish four types of mechanisms: compliance review and oversight, risk monitoring and information sharing, cross-border dispute mediation and arbitration, and mutual recognition of standards and regulatory alignment.
    • Practicing multilateralism (Chapter 05): integrate governance resources with the United Nations at the core; recommend that an independent international scientific panel lead the establishment of a global AI science communication system and introduce a risk-based tiered and categorized dialogue mechanism; explore establishing, under the UN framework, a global AI governance platform with coordination, deliberation, standard-setting, and necessary implementation oversight functions.

    Summary of recommendations in the appendix

    • The appendix lists in table form, across four dimensions, the key issues, recommended mechanisms, expected outcomes, and corresponding international experience references (such as the IPCC data-sharing platform, IAEA verification mechanism, WTO dispute settlement mechanism, WHO IHR Emergency Committee, etc.).
  4. What experiments did the paper conduct?

    The report has no experiments; its content is a review of governance practices and policy recommendations, citing multiple third-party statistical data to describe the current state of AI risks and governance participation.

    This report contains no experiments, model evaluations, or quantitative validation; its content is a qualitative review of policy documents and governance practices, citations of third-party statistical data, and mechanism design recommendations. The following summarizes the main data cited in the report and their sources.

    Risk and governance data cited in the report

    DataValueSource (source cited by the report)
    Number of global AI risk incidents (2019→2024)about 400 → more than 7,900Figure 1, data from OECD AI Incidents Monitor
    Share of incidents in the robustness and digital security, human rights and privacy, and transparency and accountability categoriesmore than 60%Figure 1
    Attack incidents recorded by ENISA (2023.7–2024.6)11,079, including 322 cross-border attacksSection 01(a), ENISA Threat Landscape 2024
    Share of main attacked industriespublic administration about 19%, transport 11%, finance 9%, digital infrastructure 8%Section 01(a)
    Year-on-year growth of generative AI-driven deepfake activity (2024)118%Section 01(a)
    Projected annual global economic loss from related cybercrime (2025)USD 10.5 trillionSection 01(a), Entrust report
    Frequency of deepfake attacks in 2024on average one every five minutesSection 02(b)
    Year-on-year growth of AI-assisted digital document forgery (2024)244%Section 02(b)

    Development gap and governance participation data cited in the report

    DataValueSource
    2024 US private-sector AI investmentabout USD 10.91 billion, which the report states is nearly 12 times that of China (USD 930 million)Section 03(a), AI Index Report 2025
    Concentration of global corporate R&D spendingabout 40% concentrated in 100 large technology companiesSection 03(a), UNCTAD report
    Share of the world's most powerful supercomputers owned by Global North countries75%, with Africa as a whole less than 1%Section 03(b)
    Relative cost for African countries to obtain core computing equipment such as GPUsthe report states it is 10 to 30 times that of relatively developed countries in per capita economic termsSection 03(b)
    UN member states participating in all seven major AI governance initiativesonly 7 of 193 member states, with 118 completely absentSection 03(c), final report of 《Governing AI for Humanity》
    Countries that signed all five key international AI safety declarationsonly 6 of 40 surveyed countriesSection 03(c), Global Index for AI Safety 2025

    Notes

    • The above figures are all third-party statistics cited by the report; the report itself did not conduct data collection or statistical analysis; Figure 1 is a chart of risk incident trends and type distribution, and in the converted text only the figure caption and the values restated in the body text can be seen.
    • The report does not report any models, benchmarks, metrics, or review settings.
  5. What are some points that can be further explored?

    The report has no limitations discussion section; its content is a qualitative policy review and recommendations, and it contains no experiments or quantitative validation.

    Limitations and future directions noted by the authors

    • The report does not include Limitations, Discussion, or Future Work sections and does not specifically discuss its own limitations.
    • In Section 05(b), the report proposes forward-looking directions: further explore, under the UN framework, establishing a global AI governance platform with stronger coordination and implementation capacity; rely on UN General Assembly mandates and institutions such as the UN Office for Digital and Emerging Technologies (ODET) to establish a cross-agency agenda coordination system; and gradually establish member-state compliance obligations through instruments such as international conventions. These are policy recommendations proposed by the report rather than completed matters.

    Scope of experimental coverage

    • The report conducted no experiments, model evaluations, or benchmark tests; all content consists of governance practice reviews, citations of third-party data, and mechanism design recommendations.
    • The cited statistical data are all quoted from third-party sources (OECD AI Incidents Monitor, ENISA, AI Index Report 2025, UNCTAD, Global Index for AI Safety 2025, etc.); the report does not report data it collected or verified itself.
    • The governance practice review covers the UN system, about 10-plus countries and regional organizations, multiple international organizations and summit mechanisms, and self-regulatory frameworks of companies such as Anthropic, OpenAI, Google DeepMind, and Meta (Section 01(b)).
    • The comparative analysis covers multilateral governance frameworks in five domains: nuclear safety, climate change, international trade, public health, and aviation and maritime (Table 1).
    • The report does not state the selection criteria, search methods, or time range for the literature review.
  6. Summarize the main content of the paper

    The World Internet Conference report diagnoses the fragmentation of global AI governance and proposes recommendations for a UN-centered global AI safety governance framework organized around four pillars.

    This is a policy report released in 2025 by the Security and Governance Program of the World Internet Conference Artificial Intelligence Committee, on the topic of mechanism design for a global AI safety governance framework.

    • Problem: AI risks spread across borders; from 2019 to 2024, global AI risk incidents increased from about 400 to more than 7,900 (Figure 1), while the global governance system is fragmented, regulation by a single country can hardly constrain cross-border risks, and geopolitical factors further weaken the space for cooperation.
    • Diagnosis: The report summarizes the key challenges into four points—technological uncertainty leads to governance lag, development gaps among countries lead to divergent demands, responsibilities of multiple stakeholders are unclear, and geopolitics leads to division (Section 01(c)).
    • Method: It systematically reviews governance practices of the United Nations, countries, regions, international organizations, summits, and academia and industry, and draws on multilateral governance experience from five domains—IAEA, IPCC, WTO, WHO, ICAO/IMO (Table 1)—while pointing out that the characteristics of AI mean existing institutions cannot simply be copied.
    • Core recommendations: It proposes mechanism design around the four pillars of "security, inclusion, responsibility, multilateralism," including cross-national technical monitoring and risk early warning, international consensus on risk classification and grading, cross-border content traceability, joint evaluation of frontier models, governance of key resources, emergency intervention mechanisms, differentiated responsibility sharing, compliance review and dispute arbitration, and integrating governance resources with the United Nations at the core and exploring the establishment of a global governance platform (Chapters 02–05, appendix).
    • Governance participation data cited: only 7 of 193 UN member states participate in all seven major AI governance initiatives, with 118 completely absent; only 6 of 40 surveyed countries signed all five key international AI safety declarations (Section 03(c)).
    • Authors' conclusion: The report advocates a UN-centered approach, in accordance with the principles of sovereign equality, international rule of law, and multilateralism, to jointly build a global AI governance framework that balances security, development, fairness, and inclusion.
Read originalarxiv.org