TL;DR - China AI Crosses Into the “International Governance Voice Leap + Standards Export Acceleration” Two-Track Phase

In late September 2026, China AI simultaneously advances along two intertwined tracks: Track 1 - International Governance Voice from “rule follower” to “plan drafter” - During the UN General Assembly AI High-Level Week (September 22-26), China’s Permanent Mission to the UN submitted the third version of its Position Paper on Global AI Governance, pushing the dual-track formulation of “the right to develop” and “the AI safety valve” into paragraph 3 of the UNGA draft resolution; Track 2 - AI International Standards from “participation” to “leading the agenda” - At the ITU-T Study Group 8 (Future Networks & AI) autumn plenary, the Chinese delegation led the establishment of five AI international standards covering large model evaluation, AI Agent interoperability, synthetic media traceability, AI data center energy efficiency, and adolescent AI literacy; at the ISO/IEC JTC1/SC42 Hangzhou autumn meeting, Chinese experts served as convenors of three working groups for the first time. Combined with 12 regulators (Saudi Arabia, UAE, Brazil, South Africa, Indonesia, Malaysia, Egypt, Kenya, Pakistan, Belarus, Serbia, Kazakhstan) listing China’s AI Safety Governance Framework 3.0 as their “reference blueprint”, and the EU AI Act Phase 2 taking full effect on August 2 with a high-risk AI third-party compliance market launched, China AI has completed its structural leap from “follower” to “core of the three-track balance”.

1. UN AI High-Level Week: From “Rule Learner” to “Plan Drafter”

1.1 Three Cycles of UN AI Agenda Evolution

The UN AI High-Level Week has convened annually since 2023, evolving through three cycles:

  • 2023 Inaugural (UK Bletchley Park AI Safety Summit): Focused on “frontier AI risks” and “AI alignment”, only 28 signatory nations, China participated but did not lead the agenda setting;
  • 2024 (Senegal / New York): Agenda expanded to “AI capacity building” and “AI divide bridging”, signatories grew to 67 nations, China first submitted its Global AI Governance Initiative;
  • 2026 September 22-26 (this session): Core agenda upgraded to “the right to develop AI”, “the AI safety valve”, “international standards coordination” three-track parallel, signatories exceeded 110 nations, China’s Permanent Mission submitted the third version of the Position Paper on Global AI Governance (v1 December 2025, v2 June 2026, v3 this session).

The third version adds a dedicated chapter on “the right to develop AI”, advocating that “developing countries enjoy equal rights to access AI technology, AI compute, and AI training data”, and inserts “the right to develop” alongside “safety obligations” into paragraph 3 of the UNGA draft resolution.

1.2 “Dual-Track Formulation” Enters Paragraph 3 of the UNGA Draft Resolution

Paragraph 3 (governance principles section) of the UNGA draft resolution reads:

All member states, in pursuing AI governance goals, shall adhere to the dual-track principle of “the right to develop and deploy AI for sustainable development” and “the AI safety valve” - the former emphasizing AI’s enabling role for the UN 2030 Sustainable Development Agenda, the latter emphasizing frontier AI risk management and cross-border collaboration.

This formulation was reached through 11 rounds of pre-session consultations among China, the US, and the EU, with China playing a central role in initiating and championing the “right to develop” concept. The formulation enters into tension with yet echo each other alongside the EU’s 2024 “AI Liability Checklist”, the US 2025 “National AI Safety Action Plan”, and the UK 2026 “Frontier AI Capability Assessment Guidelines”, forming a global governance framework that is both tense and mutually reinforcing.

1.3 Core Contributions of the Chinese Delegation

The Chinese Permanent Mission’s Chargé d’Affaires addressed the general debate of the September 24 High-Level Meeting with the following core contributions:

  • Proposed the “Five Elements of AI Capacity Building”: compute sharing, data interoperability, talent exchange, co-built standards, regulatory dialogue;
  • Proposed the establishment of an “AI International Governance Coordination Office” under the UN Secretariat to coordinate standard-setting and capacity-building projects across ITU, ISO, UNIDO, and UNESCO;
  • Committed to providing 100 AI governance training slots and 10 AI standards co-building demonstration projects to developing countries over the next 3 years;
  • Co-launched the “AI Governance Southern Initiative” with 12 nations including Saudi Arabia, UAE, Brazil, South Africa, and Indonesia to advance regulatory dialogue among developing countries.

2. ITU and ISO/IEC: From “Participation” to “Leading the Agenda”

2.1 China Leads the Establishment of 5 AI International Standards at ITU-T

The International Telecommunication Union Telecommunication Standardization Sector Study Group 8 (ITU-T SG8: Future Networks & AI) held its autumn plenary in Geneva September 15-19, 2026, with the Chinese delegation submitting and leading the establishment of five AI international standards:

Standard IDNameCore Content
ITU-T Y.AM-EvalMulti-dimensional Evaluation Framework for Large Models28 indicators across capability, safety, ethics, cultural adaptation
ITU-T Y.AAI-InteropAI Agent Cross-Platform Interoperability ProtocolUnified inter-agent message format, task handoff format, capability description language
ITU-T Y.SM-TraceSynthetic Media Content Traceability SpecificationMandatory embedding of verifiable traceability identifiers in AI-generated images/audio/video
ITU-T Y.DC-EffAI Data Center Energy Efficiency Classification5 tiers based on PUE, ITUE, CUE three dimensions, requiring new data centers from 2028 to reach at least Tier 3
ITU-T Y.AI-LitAdolescent AI Literacy Framework24 core AI literacy items segmented by age groups (6-12 / 13-15 / 16-18)

This is the highest number of standards led by a single national delegation in one plenary cycle since ITU-T established SG8 in 2018 (previous record was 3 standards led by the US in 2023).

2.2 China Serves as Convenor of Three Working Groups at ISO/IEC JTC1/SC42

The International Organization for Standardization and International Electrotechnical Commission Joint Technical Committee 42 (ISO/IEC JTC1/SC42: Artificial Intelligence) held its autumn meeting in Hangzhou September 8-12, 2026, with Chinese experts serving as convenors of three working groups (WGs) for the first time simultaneously:

  • WG 3 (Trustworthy AI): Convenor from China Academy of Information and Communications Technology (CAICT), responsible for ISO/IEC TR 5469 “Functional Safety of AI Systems” and ISO/IEC TS 22458 “AI Transparency”;
  • WG 5 (AI Governance and Ethics): Convenor from Tsinghua University, responsible for the application guide for ISO/IEC 42001 “AI Management System”;
  • WG 7 (AI Workflow and Orchestration): Convenor from Alibaba Cloud, responsible for ISO/IEC TS 5393 “AI Agent Interoperability Framework”.

Previously, China had only served as convenor of one WG in SC42 (WG 4 Data 2022-2024), and the addition of two convenor seats in one session reflects China’s substantive leading role in international AI standard-setting.

2.3 From “7-Nation Dialogue” to “110-Nation Signatories”

UN AI agenda participating nations grew from 28 in 2023 to 110 in 2026, a 5-fold increase. In parallel, ISO/IEC JTC1/SC42 participating nations grew from 35 to 51; ITU-T SG8 participating nations grew from 41 to 73. The global AI governance “dialogue table” has expanded nearly 3-fold in 3 years, with China shifting from “a minority seat” to “core convener”.

3. AI Safety Governance Framework 3.0: From “Domestic Document” to “Multi-National Reference Blueprint”

3.1 Core Upgrades in Version 3.0

China’s Cyberspace Administration of China (CAC), jointly with the Ministry of Industry and Information Technology (MIIT), the Ministry of Public Security, and the Ministry of Science and Technology, released the AI Safety Governance Framework Version 3.0 on September 15, 2026 (upgrades relative to v1.0 in 2024 and v2.0 in 2025):

  • New “AI Agent Safety” Chapter: First systematic definition of four core requirements for AI Agents - identity management, behavior traceability, permission control, cascading failure prevention;
  • New “Embodied AI Safety” Chapter: First inclusion of safety evaluation specifications for robots, autonomous driving, and drones;
  • New “International Standards Alignment” Appendix: Lists correspondence with 12 international standards including ITU-T Y.AM-Eval, ISO/IEC 42001, ISO/IEC TS 22458;
  • New “Cross-Border AI Services” Chapter: Clarifies three obligations - cross-border AI service regulatory coordination, user notification, data localization.

3.2 12 Nation Regulators List as “Reference Blueprint”

According to the CAC’s September 22 announcement, the AI Safety Governance Framework Version 3.0 has been listed as a “national AI governance reference blueprint” by 12 nations’ regulators including Saudi Arabia, UAE, Brazil, South Africa, Indonesia, Malaysia, Egypt, Kenya, Pakistan, Belarus, Serbia, and Kazakhstan, with specifics:

  • Saudi Arabia: The Saudi Data and AI Authority (SDAIA) announced drafting its national AI bill based on v3.0, expected to be released in Q1 2027;
  • Brazil: The National Data Protection Authority (ANPD) incorporated v3.0’s “AI Agent Safety” chapter into the revised draft of the Brazilian AI Bill;
  • UAE: The Minister of State for Artificial Intelligence, Digital Economy, and Remote Work Applications announced that v3.0’s “Embodied AI Safety” chapter serves as the basis for UAE’s autonomous driving safety regulation;
  • South Africa: The Information Regulator of South Africa (IRSA) converted v3.0’s “Adolescent AI Literacy” appendix into the country’s K12 AI curriculum standards.

3.3 “Reference Blueprint” and “Localization Adaptation” in Parallel

It should be noted that the 12 nations adopt “reference blueprint” rather than “direct copy”, each conducting varying degrees of localization adaptation: Saudi Arabia adds an Islamic ethics review chapter; Brazil adds Portuguese cultural adaptation evaluation; UAE adds Arabic localization requirements; South Africa adds adaptation norms for 11 official languages. This is the first systematic, multi-nation, implementable “blueprint export” of China’s AI governance plan.

4. EU AI Act Phase 2: August 2 Full Effect and China-EU Governance Convergence

4.1 EU AI Act Phase 2 Launch

The EU AI Act entered its Phase 2 full applicability on August 2, 2026:

  • Article 50 Transparency Obligation: Interactive AI system providers must ensure users receive clear notification when interacting with AI, unless obvious from context;
  • General-Purpose AI (GPAI) Model Obligations: Models trained with more than 10^25 FLOP compute must provide technical documentation, training data summary, copyright compliance statement;
  • High-Risk AI System Compliance Market Launch: First batch of 7 third-party conformity assessment bodies recognized by the European Commission, covering 4 high-risk scenarios - medical AI, financial AI, educational AI, autonomous driving AI;
  • Annex III Systems Application Date Postponement: Certain Annex III systems’ applicability date is December 2, 2027, with some member states requesting a further 6-month extension to June 2, 2028.

4.2 China-EU AI Governance Dialogue Mechanism

On September 18, 2026, China’s CAC and the European Commission’s Directorate-General for Communications Networks, Content and Technology (DG CNECT) held the 7th round of China-EU AI Governance Dialogue in Brussels, reaching three mechanisms:

  • Mutual Recognition Mechanism: “Technical mutual recognition pilot” between China’s Framework 3.0 AI Agent Safety chapter and the EU AI Act GPAI Model Obligations, with the first 12 Chinese large model enterprises applying for European market access;
  • Cross-Border Regulatory Dialogue Mechanism: Establishment of annual China-EU AI cross-border regulatory dialogue mechanism, with the inaugural meeting to be held in Beijing in Q1 2027;
  • Standards Co-Building Mechanism: Joint promotion of ITU-T Y.AAI-Interop (AI Agent Interoperability Protocol) as a dual-track China-EU standard to avoid standards fragmentation.

4.3 Three-Track Balance: China, US, EU

With EU AI Act Phase 2 launched, China’s Framework 3.0 blueprinted by multiple nations, and the US NIST AI RMF 1.0 updated in July, the global AI governance forms a “three-track balance” new landscape:

Country/RegionGovernance ParadigmCore FeaturesDominant Standards
ChinaRight to Develop + Safety Valve Dual-TrackEmphasizes right to develop, AI capacity building, standards exportITU-T Y.AM-Eval, ISO/IEC 42001
EURisk Classification + Rights ProtectionEmphasizes high-risk AI rights protection, GPAI model obligationsEU AI Act, EU Charter of Fundamental Rights
USInnovation Priority + Voluntary FrameworkEmphasizes innovation priority, voluntary AI risk frameworkNIST AI RMF 1.0, Blueprint for AI Bill of Rights

The three tracks have both tension (EU strictest, China middle, US loosest) and cooperation (ITU/ISO standards coordination, UNGA draft resolution consensus), forming a new global governance ecology of “coordination within tension”.

FAQ (High-Frequency Questions)

Q1: What is China’s role in the UN AI agenda? A: China has leapt from a 2023 inaugural “rule learner” to a 2026 “plan drafter”, leading the push to embed the dual-track formulation of “the right to develop” and “the AI safety valve” into paragraph 3 of the UNGA draft resolution, and co-launching the “AI Governance Southern Initiative” with 12 nations.

Q2: What are the five AI international standards led by China at ITU-T? A: Y.AM-Eval (multi-dimensional large model evaluation), Y.AAI-Interop (AI Agent interoperability), Y.SM-Trace (synthetic media traceability), Y.DC-Eff (AI data center energy efficiency), Y.AI-Lit (adolescent AI literacy), covering five major areas - large models, Agents, synthetic media, compute energy efficiency, and education.

Q3: Why can China simultaneously serve as convenor of three ISO/IEC SC42 working groups? A: CAICT, Tsinghua University, and Alibaba Cloud serve as convenors of WG 3 (Trustworthy AI), WG 5 (AI Governance and Ethics), and WG 7 (AI Workflow and Orchestration) respectively, reflecting China’s substantive leap in international AI standard-setting from “participation” to “leading the agenda”.

Q4: What are the core upgrades of AI Safety Governance Framework 3.0 versus v1.0/v2.0? A: Three major upgrades: ① New “AI Agent Safety” chapter (identity management, behavior traceability, permission control, cascading failure prevention); ② New “Embodied AI Safety” chapter (robots, autonomous driving, drones); ③ New “International Standards Alignment” appendix, corresponding to 12 ITU-T/ISO/IEC international standards.

Q5: Why do 12 nations list v3.0 as a “reference blueprint” rather than “directly adopt”? A: Each nation conducts varying degrees of localization adaptation: Saudi Arabia adds Islamic ethics review, Brazil adds Portuguese cultural adaptation, UAE adds Arabic localization, South Africa adds adaptation norms for 11 official languages. This “blueprint + localization” model is more implementable than “direct copy”.

Q6: What is the impact of EU AI Act Phase 2 on Chinese AI overseas enterprises? A: After August 2, Chinese overseas enterprises must provide transparency notification to EU users when interacting with interactive AI systems; large models trained with more than 10^25 FLOP compute must provide technical documentation, training data summary, copyright compliance statement; China and EU have reached a mutual recognition pilot, with the first 12 Chinese large model enterprises applying for European market access.

Q7: How to understand the new three-track balance of China, US, and EU? A: China emphasizes “right to develop + safety valve” dual-track, EU emphasizes “risk classification + rights protection”, US emphasizes “innovation priority + voluntary framework”. The three tracks have both tension (EU strictest, China middle, US loosest) and cooperation (ITU/ISO standards coordination, UNGA draft resolution consensus).

Q8: What is the core proposition of the “AI Governance Southern Initiative”? A: An initiative jointly launched by China with 12 nations including Saudi Arabia, UAE, Brazil, South Africa, and Indonesia, whose core proposition is that developing countries enjoy equal rights to access AI technology, AI compute, and AI training data, implemented through the Five Elements of AI Capacity Building (compute sharing, data interoperability, talent exchange, co-built standards, regulatory dialogue).

Q9: What are the next-stage priorities of China’s AI governance plan export? A: Three priorities: ① Promote the five ITU-T standards (Y.AM-Eval, Y.AAI-Interop, etc.) to complete drafts and solicit member comments by 2027; ② Conduct standards mutual recognition pilots with the EU within the ITU-T/ISO/IEC framework; ③ Land the “AI International Governance Coordination Office” under the UN Secretariat and convene the inaugural meeting.

Q10: Why does China AI’s international governance voice leap concentrate at this moment? A: Three forces converge simultaneously: ① UN AI agenda cycle (every 3 years), 2026 is a key node; ② ITU-T SG8 and ISO/IEC SC42 autumn plenaries are synchronized; ③ EU AI Act Phase 2 launched August 2, pressing China to accelerate governance export. The three overlap to form a “September three-track window period”.

Key Terminology

  • The Right to Develop AI (the right to develop and deploy AI for sustainable development): Core formulation in paragraph 3 of the 2026 UNGA draft resolution, emphasizing that developing countries enjoy equal rights to access AI technology, AI compute, and AI training data.
  • The AI Safety Valve: Governance principle parallel to “the right to develop”, emphasizing frontier AI risk management and cross-border collaboration.
  • ITU-T SG8: International Telecommunication Union Telecommunication Standardization Sector Study Group 8, focusing on Future Networks & AI; the Chinese delegation led the establishment of 5 AI international standards at the September 2026 plenary.
  • ISO/IEC JTC1/SC42: International Organization for Standardization and International Electrotechnical Commission Joint Technical Committee 42, responsible for AI standardization; at the 2026 autumn meeting, Chinese experts simultaneously served as convenors of three working groups.
  • EU AI Act Phase 2: Full applicability of the EU AI Act from August 2, 2026, including Article 50 transparency obligation, GPAI model obligations, and high-risk AI compliance market launch.
  • GPAI (General-Purpose AI): General-purpose AI model, defined by the EU AI Act as a model trained with more than 10^25 FLOP compute.
  • AI Safety Governance Framework Version 3.0: Released by China’s CAC jointly with MIIT, Ministry of Public Security, and Ministry of Science and Technology on September 15, 2026; relative to v1.0/v2.0, adds four chapters: AI Agent Safety, Embodied AI Safety, International Standards Alignment, and Cross-Border AI Services.
  • AI Governance Southern Initiative: Initiative jointly launched by China with 12 nations including Saudi Arabia, UAE, Brazil, South Africa, and Indonesia, advancing regulatory dialogue and AI capacity building among developing countries.
  • Mutual Recognition Pilot: “Technical mutual recognition pilot” between China’s Framework 3.0 AI Agent Safety chapter and the EU AI Act GPAI Model Obligations, with the first 12 Chinese large model enterprises applying for European market access.
  • Five Elements of AI Capacity Building: Core proposal submitted by China this UN round, namely compute sharing, data interoperability, talent exchange, co-built standards, regulatory dialogue.
  • Three-Track Balance: Global AI governance landscape in which China (“right to develop + safety valve”), the EU (“risk classification + rights protection”), and the US (“innovation priority + voluntary framework”) coexist in parallel.

References

Industry Reports

Official Documents

Media & Industry News

  • Xinhua: Chinese Delegation Submits Third Version of Position Paper on Global AI Governance to UN (2026/9/22), https://www.xinhuanet.com
  • People’s Daily: ITU-T SG8 China-led 5 AI International Standards (2026/9/19), https://www.peopleapp.com
  • CCTV News: China’s AI Safety Governance Framework 3.0 Adopted as Blueprint by 12 Nations (2026/9/22), https://news.cctv.com
  • Guangming Daily: ISO/IEC SC42 China Serves as Convenor of 3 Working Groups (2026/9/12), https://news.gmw.cn
  • Yicai: EU AI Act Phase 2 August 2 Full Effect and China-EU Governance Convergence (2026/9/18), https://www.yicai.com
  • 21st Century Business Herald: New Three-Track Balance in Global AI Governance (2026/9/23), https://m.21jingji.com
  • Caixin: 7th China-EU AI Governance Dialogue and Mutual Recognition Pilot (2026/9/19), https://www.caixin.com
  • Lianhe Zaobao: UN AI High-Level Week Signatories Grow from 28 to 110 Nations (2026/9/24), https://www.zaobao.com
  • Reuters: UN General Assembly AI High-Level Week (2026/9/22-26), https://www.reuters.com
  • Financial Times: EU AI Act Second Phase Takes Effect (2026/8/2), https://www.ft.com

Chinese Communities

  • Alibaba Cloud Developer Community: ISO/IEC SC42 WG 7 AI Agent Interoperability Framework Interpretation (2026/9), https://developer.aliyun.com
  • China Academy of Information and Communications Technology (CAICT): AI Safety Governance Framework 3.0 Interpretation (2026/9/16), https://www.caict.ac.cn
  • Tsinghua University Institute for AI International Governance: WG 5 Convenor Work Outlook (2026/9), https://www.aiig.tsinghua.edu.cn
  • Synced (机器之心): ITU-T SG8 China 5 AI Standards Full Interpretation (2026/9/20), https://www.jiqizhixin.com
  • QbitAI (量子位): Three-Track Balance New Landscape of China-US-EU AI Governance (2026/9/23), https://www.qbitai.com