State of AI Safety in China (2025) Report released
Concordia AI is proud to have released the State of AI Safety in China (2025) report at the World AI Conference (WAIC) in Shanghai! You can download the full report here and on our website.
First published in 2023 and updated in 2024, this annual report offers one of the most comprehensive overviews of how China is approaching the safety and governance of general-purpose AI systems.
The new edition spans updates from May 2024 to June 2025. Boasting around 100 pages, it is packed with new insights. While some findings will be familiar to regular readers of our newsletter, the report includes extensive new content, not previously published, such as:
A systematic analysis of China’s AI safety standards, and where they are headed;
A quantitative assessment of frontier AI safety research by Chinese scholars;
A systematic mapping of safety-related discourse at major Chinese AI conferences;
A first-of-its-kind analysis of AI safety disclosures in technical model cards by major Chinese AI developers;
And much more.
Key findings
Domestic Governance
Domestic rhetoric and policy include increasingly prominent and specific calls for AI risk mitigation, while continuing to emphasize the complementarity between AI safety and development. At the Third Plenum, one of China’s most important political events in 2024, AI safety was formally elevated to a national priority. In April 2025, the country’s seniormost officials in the Communist Party of China (CPC) Politburo held a study session dedicated to AI, calling for monitoring, early warning, and emergency response systems to ensure AI safety, reliability, and controllability. This marked a significant endorsement of AI risk mitigation. AI safety has also gained prominence in official meetings and emergency planning documents related to national security and public safety, suggesting growing recognition of AI’s potential for large-scale risks.
China is implementing its AI regulations through an expanding AI standards system. While a comprehensive national AI Law remains unlikely in the near future, China has issued detailed regulations on AI-generated content labeling and watermarking. It also continues to implement existing regulations, which require pre-deployment registration and safety testing of certain types of AI models. These mandates are operationalized through a rapidly expanding system of standards: from January to May 2025 alone, China issued as many national AI standards as in the preceding three years combined. Although current regulations and standards primarily address near-term risks, official standard-setting plans reveal intentions to address severe risks such as AI’s impact on cybersecurity or loss of control.
International Governance
China has emphasized AI safety and global AI capacity-building as key themes in its international AI diplomacy. At Davos in January 2025, China’s top science and technology official warned that developing AI without safety measures is like driving on a highway without brakes, and cautioned against unchecked international competition. At the Paris AI Action Summit, leading Chinese academic and policy institutions launched the “China AI Safety & Development Association” (CnAISDA), positioning itself as China’s counterpart to other national AI safety institutes. Meanwhile, the Chinese government has become more proactive on AI capacity-building in the Global South by sponsoring a UN resolution (adopted by over 140 countries), launching an “Action Plan,” and co-founding a “Group of Friends” on AI capacity-building.
China has launched bilateral AI dialogues with several key countries, though the outlook for these engagements remains mixed. While Chinese President Xi Jinping and then-United States (US) President Joe Biden agreed on the importance of human control over nuclear systems in November, their intergovernmental AI dialogue has not met since May 2024, and its future faces major uncertainty under the Trump administration. Meanwhile, China and the United Kingdom (UK) inaugurated a new bilateral AI dialogue in May 2025, suggesting China's continued interest in engaging willing partners.
Technical Safety Research
Frontier AI safety research output in China has expanded rapidly. From June 2024 to May 2025, Chinese scholars published ~26 papers/month, more than double the previous year’s output. Previously underexplored areas, such as alignment of superhuman systems and mechanistic interpretability, have become popular topics. Research on deception, unlearning, and CBRN (chemical, biological, radiological and nuclear) misuse has also expanded, though the latter remains mostly benchmark-driven. Chinese researchers are taking concrete actions to address severe AI safety risks.
Expert Views on AI Safety and Governance
Expert discourse in China is placing greater emphasis on AI safety and governance. From 2023 to 2024, coverage of AI safety and governance at two major AI conferences—the World AI Conference (WAIC) and World Internet Conference (WIC)—more than doubled. In addition, WAIC was upgraded in 2024 to a “High-Level Meeting on Global AI Governance.” For the first time in the conference's history, a top central government official, Premier LI Qiang (李强), delivered the opening remarks. Notably, he highlighted AI safety and governance issues. Meanwhile, expert speeches also emphasized these topics more than in previous years.
Experts are increasingly publishing in-depth analyses of AI risks in biosecurity, cybersecurity, and open source AI. Multiple state-affiliated and independent experts have written detailed analyses on AI misuse in biosecurity and cybersecurity. Meanwhile, Chinese experts highlight the safety benefits of open source AI, while also increasingly exploring potential downstream misuse risks.
Industry Governance
Most leading Chinese foundation model developers have signed voluntary “AI Safety Commitments.” The commitments, drafted by the AI Industry Alliance of China (AIIA) in December 2024, pledge safety measures across the AI development lifecycle, including dedicated safety teams, red teaming, data security, infrastructure protection, transparency, and investment in frontier safety research.
Chinese AI developers typically implement well-known safety methods, but provide limited transparency on safety evaluations. Our survey of frontier model releases shows that nine out of thirteen developers have published technical model release cards, many reporting standard safety techniques like data filtering, reinforcement learning from human feedback (RLHF), constitutional AI, and red teaming. However, transparency remains limited: only three of the surveyed companies have published risk evaluation results.
We hope this report helps foster informed and nuanced discussions about the trajectory of AI safety and governance in China. We welcome your critical feedback and comments!