Voices

Public statements from China's AI governance ecosystem, organized by topic

Understanding how to engage with Chinese AI governance actors requires knowing how they frame key issues publicly. This page curates notable statements from official speeches, government releases, and institutional publications — presented in Chinese with English translations. Statements are time-stamped and sourced; readers are encouraged to consult the originals for full context.

This is a curated selection of public statements, not a comprehensive survey of Chinese AI discourse. Curation involves editorial choices — different curators would select different quotes. Public statements reflect strategic communication, not necessarily private positions or operational intent. Readers are encouraged to consult original sources and triangulate with other perspectives.

AI Risk Governance

As AI systems move from passive tools to autonomous agents, how should risk be categorized and responsibility allocated? Chinese scholars and regulators are developing distinctive frameworks — including a three-tier risk taxonomy, 'agile governance' through rapid regulatory iteration, granular authorization models for AI agents, and investments in mechanistic interpretability. These framings directly shape what cooperation on AI safety is possible.

Liang Zheng (梁正)

Deputy Director, Tsinghua I-AIIG; Director, AI Governance Research Center; Professor, School of Public Administration

2 quotes

Mechanistically, AI risk comes from three sources. The first is endogenous risk — risk arising from the inherent characteristics of the technical system itself, such as loss of control. The second is secondary risk at the application level — misuse, abuse, or deployment in inappropriate contexts. The third is derivative risk — broadly societal impacts like cognitive dependency and job displacement that are not directly caused by the technology itself but by how we structure our relationship with it.

Original:

从机理上看，人工智能的风险主要来自3个方面。第一类是内生性风险，即技术系统自身特性导致的风险，比如失控。第二类是应用层面的次生风险。第三类是衍生风险，与技术本身及具体用途关系不大，却带来广泛社会影响。

人民日报海外版｜梁正等：当技术挑战伦理，治理如何跟上 (Tsinghua I-AIIG WeChat) · People's Daily Overseas Edition interviewJuly 2026

Agile governance means adapting to changing conditions and adjusting in real time. For AI governance, we have always emphasized 'small steps, fast pace.' When new problems emerge, don't lock everything down immediately — first assess whether existing governance tools are sufficient.

Original:

敏捷治理强调因时而变，并根据形势及时调整。在人工智能治理上，我们一直强调“小步快走”。新问题出现时，不要一上来就管死，而应先研判现有治理工具是否够用。

人民日报海外版｜梁正等：当技术挑战伦理，治理如何跟上 (Tsinghua I-AIIG WeChat) · People's Daily Overseas Edition interviewJuly 2026

Zhang Linghan (张凌寒)

Director, AI Law Research Institute, China University of Political Science and Law; Expert, UN AI Advisory Body

2 quotes

The primary standard for liability should be whether the agent's behavior falls within the scope of genuine, clear, and continuous human authorization. Human authorization of an AI agent cannot simply be equated with a one-time blanket consent. It must distinguish between specific scenarios, behavior types, and risk levels — is this limited authorization? Temporary? General? Continuous?

Original:

责任认定的首要标准，是看智能体的行为是否处于人类真实、清晰、持续的授权范围内。人类对智能体的授权，不能简单等同于一次性概括同意，而应区分具体场景、行为类型与风险等级——是有限授权、临时授权，还是概括性授权、持续性授权？

算法成“看不见的老板”、AI“闯祸”责任归属不明，学者呼吁→ (Pengpai News) · InterviewJuly 17, 2026

When an AI agent exceeds its authority, we cannot only hold the user accountable. We must also trace back to whether the model provider, the platform deployer, and the application developer fulfilled their duty of care — including technical controls and risk warnings. The future AI liability system will not be one actor bearing one responsibility, but multiple actors bearing distributed responsibility.

Original:

当智能体越权，不能只追问使用者，还要回溯模型提供者、平台部署方、应用开发方是否尽到了技术管控、风险提示等注意义务。未来的人工智能责任体系，并非单一主体的单一责任，而是多元主体的责任。

算法成“看不见的老板”、AI“闯祸”责任归属不明，学者呼吁→ (Pengpai News) · InterviewJuly 17, 2026

Tang Jie (唐杰)

Founder, Zhipu AI

2 quotes

When the most advanced frontier models overseas are delayed from full public release due to risk considerations, and their leaders publicly warn that AI's profound impact will reshape global power structures, we should be even more clear-headed: the pursuit of superintelligence and the research on superalignment must advance in lockstep.

Original:

当海外最前沿的顶级模型因风险考量而暂缓全面公开发布，其企业负责人公开警示AI的深远影响将深刻重塑全球力量格局时，我们更应清醒：超级智能的实现与超级对齐的研究，必须同步推进。

晚点独家｜唐杰内部信 (LatePost) · Internal letterJuly 2026

The more powerful the capability, the more robust the safety constraints must be. From its founding, Zhipu established a principle: AI must serve human welfare and national strategy. The company rejects bolt-on safety patches, instead writing human ethics, social norms, and national laws into the model's value function as foundational axioms. Zhipu plans to invest tens of billions in mechanistic interpretability research.

Original:

能力越强大，安全约束机制也必须越稳固。智谱创立伊始即确立准则：AI必须服务于人类福祉，服务于国家战略。公司摒弃外挂式安全补丁，坚持将人类伦理、社会规范及国家法律法规作为底层公理写入模型价值函数；计划投入百亿级资源攻坚“机械可解释性”。

晚点独家｜唐杰内部信 (LatePost) · Internal letterJuly 2026

Xiao Qian (肖茜)

Deputy Director, Tsinghua Institute for AI International Governance

Future global AI governance must not only manage risks but also create development opportunities; it must not only prevent technological loss of control but also prevent development imbalances.

Original:

未来全球AI治理不仅要管控风险，更要创造发展机会；不仅要防范技术失控，更要防止发展失衡。

新华社·肖茜｜未来全球AI治理要管控风险，更要创造机会 (Tsinghua I-AIIG WeChat) · State media interviewJuly 17, 2026

Zhou Bowen (周伯文)

Director and Chief Scientist, Shanghai AI Laboratory

We have built a 'security-as-a-service' operating system and developed an intrinsic autonomous safety architecture for AI (R²AI — Resilient & Responsible AI), raising defense success rates to over 42%.

Original:

我们打造了“安全即服务”的书安操作系统，研发了人工智能的内生自主安全架构（R²AI），把防御成功率提升到了42%以上。

周伯文：AGI的下一程，迎接科学元认知时刻 | WAIC 2026 (Shanghai AI Lab WeChat) · KeynoteJuly 17, 2026

Zhuang Rongwen (庄荣文)

Director, Cyberspace Administration of China (CAC); Deputy Head, Central Propaganda Department

Promote healthy technological development. Balance development and security, encourage new technology development in the cyberspace domain, refine the tiered and classified safety regulatory mechanism, strengthen compliance guidance, combat technology abuse, strengthen AI safety governance, improve the filing management system for large model services, and ensure that new technologies and applications remain safe, controllable, and develop toward the good.

Original:

推动技术健康发展。统筹发展和安全，鼓励网信领域新技术发展，积极运用新技术赋能内容建设和乱象治理，完善分级分类的安全监管机制，加强合规指导，打击技术滥用，强化人工智能安全治理，完善大模型服务备案管理，确保新技术新应用安全可控、向善发展，筑牢网络安全和数据安全防线。

庄荣文：提高网络生态治理效能营造风清气正网络空间 (CAC WeChat / Qiushi journal) · Signed articleJuly 31, 2026

Xue Lan (薛澜)

State Council Counselor; Dean of Schwarzman College; Director, Institute for AI International Governance, Tsinghua University

From a technical perspective, AI has advanced at a breathtaking pace over the past two years. By comparison, the construction of socio-technical systems has lagged far behind. We can now say that AI development has entered the 'second half' — and building the socio-technical systems needed for AI's application in society has become extremely urgent.

Original:

从人工智能的技术层面看，过去两年的发展是日新月异。与此相比，社会技术系统的构建更加迟缓。现在可以说人工智能发展进入到下半场了，构建人工智能技术在社会中应用的社会技术系统就变得非常急迫了。

薛澜：《人工智能技术的社会应用——治理挑战》 (Tsinghua I-AIIG WeChat) · KeynoteJuly 3, 2026

Zeng Yi (曾毅)

Director, Beijing Institute of AI Safety and Governance; Member, UN High-Level Advisory Body on AI

2 quotes

The evolution of frontier AI systems has, in a meaningful sense, outpaced human understanding. Effective control is stretched thin. We are no longer dealing with tools that merely execute instructions. We are building intelligent systems capable of reasoning, planning, and generating novel solutions across domains — from scientific discovery to economic decision-making — sometimes producing outputs that exceed the tasks we assigned them.

Original:

前沿人工智能系统的演化速度，从某种意义而言已经超越了现阶段人类的理解，有效的管控已捉襟见肘。我们面对的不再是只会执行指令的工具。我们正在打造能够推理、规划、跨领域生成全新解决方案的智能系统——从助力科学发现到支撑经济决策，其产出有时甚至超出我们赋予人工智能的任务预期。

曾毅教授在CGTN“智能文明”特别节目 (Beijing AISI WeChat) · CGTN keynoteJuly 17, 2026

Third is the competitive dynamics effect. When countries and companies race to develop more powerful systems, ethics and safety are often relegated to afterthought add-ons, or even viewed as obstacles to progress. This is a dangerous fallacy. Ethics and safety are not merely brakes to be applied when risks emerge — they are the steering wheel that guides the direction of development.

Original:

第三是竞争博弈效应。当各国、各个企业争相研发更强性能的智能系统时，伦理与安全往往沦为事后补救的附加项，甚至是被视作发展的阻力。这是一种危险的谬误。伦理与安全并不只是在风险出现以及必要时发挥刹车的作用，而更是指引发展方向的方向盘。

曾毅教授在CGTN“智能文明”特别节目 (Beijing AISI WeChat) · CGTN keynoteJuly 17, 2026