Is China Serious About AI Safety?
China’s new AI safety body brings together leading experts — but faces obstacles to turning ambition into influence.
This summer, the World AI Conference (WAIC) in Shanghai began to live up to its name. Previously an almost exclusively domestic event, this year’s event attracted a larger group of international visitors to witness the would-be marvels of China’s AI ecosystem. It also provided an opportunity to engage foreign counterparts for one of the newest elements of that ecosystem: the China AI Safety and Development Association (CnAISDA).
Launched in February 2025 on the sidelines of the Paris AI Action Summit, CnAISDA places China among a small number of jurisdictions with dedicated AI safety institutes, or AISIs — although they increasingly go by other names. AISIs are government-backed institutions with a focus on AI risks, sometimes explicitly including catastrophic risk. Given the otherwise zygotic state of efforts to address potential catastrophic risks of frontier AI systems in China, CnAISDA is potentially a kernel of important things to come.
But does CnAISDA signal a serious national commitment to AI safety — or is it mostly talk?
To understand whether CnAISDA represents real momentum or mere symbolism, it helps to first look at how China’s domestic AI safety ecosystem has evolved over the past year, as domestic developers and researchers have begun adapting some governance measures pioneered abroad.
Several Chinese developers have signed voluntary safety commitments that largely mirror global industry promises. In December 2024, China's Artificial Intelligence Industry Alliance, a part of CAICT, released safety commitments signed by 17 major Chinese AI companies, including DeepSeek, Alibaba, and Tencent. Notably, these commitments closely mirror the Frontier AI Safety Commitments made by AI developers around the world at the Seoul AI Summit. At WAIC this year, the CAICT commitments were reiterated under the CnAISDA umbrella and slightly expanded to put marginally more emphasis on international cooperation and prevention of frontier risks.
However, some key stakeholders, notably including DeepSeek, were conspicuously absent from the second round of commitments. A key indicator for safety concerns in China will be whether Chinese developers comply with their commitments. Global compliance with the Seoul commitments is patchy, and none of the Chinese companies that signed them have fulfilled them.
A key state-backed lab has published its own risk mitigation strategy along with results from frontier AI risk evaluations that build on global best practices. Concurrent with this year’s WAIC, Shanghai AI Lab, in partnership with Concordia AI, published a Frontier AI Risk Management Framework laying out a detailed process for monitoring and managing risks from frontier AI systems. These same organizations published an analysis of the frontier risks posed by various Chinese and non-Chinese AI models, based on the Shanghai AI Lab framework.
This represented a significant step forward for the Chinese AI risk evaluations landscape, though it remains far more nascent than technical alignment, safeguards, and evaluations work in the US and UK. Previously, China had limited engagement with risks such as strategic deception in autonomous R&D and CBRN proliferation. As Anthropic co-founder Jack Clark noted, Shanghai AI Lab’s findings highlighted closer convergence with risk evaluations conducted in the US and UK, in both risks studied and methods used. Shanghai AI Lab even employed UK AISI’s Inspect evaluation software.
CnAISDA’s leadership includes some of the preeminent figures in Chinese AI policy. Fu Ying, a former vice minister of foreign affairs and one of China's most prominent diplomats, has served as the organization's primary public spokesperson. Andrew Yao, China's sole Turing Award winner, lends scientific credibility. Xue Lan, a key adviser to China's powerful State Council, provides policy expertise and bridges technical and policy communities. Other leaders featured on CnAISDA’s website include Zhou Bowen of the Shanghai Artificial Intelligence Laboratory (SHLAB), Wei Kai of the China Academy of Information and Communications Technology (CAICT), Xu Wei of Tsinghua’s Institute for Interdisciplinary Information Sciences, and Zeng Yi from the Chinese Academy of Sciences.
/inline-pitch-cta
Unlike its US and UK counterparts, which were created as new intra-government agencies, CnAISDA is a coalition of existing institutions. The most fundamental feature of CnAISDA is its decentralized structure, which follows the approach taken by the Canadian AISI and France’s INESIA. Rather than creating a new bureaucratic entity, China has assembled existing centers of expertise on AI under one umbrella. CnAISDA’s name in Mandarin emphasizes this aspect; the literal translation is not “association” but “network.”
Tsinghua University serves as the organizational hub: CnAISDA's official address and phone number are both at the university, and many of the leading staff have affiliations there. Arguably China’s most prestigious university, Tsinghua hosts important centers of activity on safety and governance, and is a top producer of safety-related research. The grouping also incorporates other leading technical organizations in AI, including SHLAB, the Beijing Academy of Artificial Intelligence (BAAI), and CAICT. SHLAB and CAICT evaluate AI systems for risk-related properties, among other functions. BAAI is a key convener and hub for much activity related to technical AI development in the capital.
The coalition design reflects political pragmatism. Rather than create a single centralized agency — which would require picking winners among competing institutions — China has opted for a coalition model. This approach allows the government to elevate multiple key voices in the Chinese AI safety conversation without having to choose among competing institutions with overlapping claims to expertise.
Limited resources. Whatever CnAISDA’s ambitions, the group faces significant hurdles as it gets established. The organization currently lacks dedicated staff, a budget of its own, or even a Chinese-language website, raising questions about its operational capacity. The nature of its relationship with the Chinese government remains ambiguous, though an official from the powerful National Development and Reform Commission has publicly confirmed CnAISDA’s government backing.
At a time when leading figures in US AI policy have signaled skepticism about global AI governance efforts, Chinese leaders have continued to emphasize support for inclusive AI development and multilateral engagement.
CnAISDA is focused primarily on international engagement, at least for now. When China’s 2024 Third Plenum decision document included a call for establishing an “AI safety supervision system,” many observers saw it as teasing the creation of a Chinese AISI, which would carry out domestic functions like testing and evaluations of frontier AI models. What has been created so far, however, is more of an engagement platform than a supervision system. CnAISDA is designed primarily as China's voice in global AI governance discussions. It is not yet set up to conduct significant independent work, whether technical research, evaluations, or policy functions. To the extent that CnAISDA takes on a more domestic role over time, this would be even stronger evidence of focus from Chinese leadership on risks from frontier AI.
/odw-inline-subscribe-cta
This international orientation was evident at CnAISDA's launch event in Paris. The organization's founding document calls for building "an inclusive global AI safety governance framework" that combines UN engagement with coordination among AI safety institutes worldwide. It advocates for international cooperation to prevent AI misuse by terrorist organizations and establishes "early warning thresholds for AI systems that may pose catastrophic or existential risks to humans."
CnAISDA’s focus on international engagement aligns with China’s other efforts on global AI governance. China’s “Global AI Governance Action Plan,” which offers a competing vision to the Trump administration’s AI Action Plan, stresses the importance of international collaboration and broadly distributed capacity-building. CnAISDA partially substantiated these ambitions by hosting several well-attended panels with international experts on topics including “The Harmonious Future of AI and Humanity” and “The Global Governance of AI Development and Safety.” Although only words, these events did mark notable discussion by Chinese and foreign experts of cutting-edge topics that could be crucial elements of future international governance, such as technical verification and systemic risks from AI agents. Many in the audience were likely exposed to these ideas for the first time.
However, China’s international approach increasingly centers on the global diffusion of AI. In his speech launching China’s Plan, Chinese Premier Li Qiang announced the Chinese government’s support for the creation of a “World AI Cooperation Organization,” provisionally to be based in Shanghai. In addition, Li announced a new AI+ International Cooperation Initiative at the high-level meeting on the Global Development Initiative in New York. That effort seeks to expand China’s efforts to diffuse AI into the “real economy” today.
These efforts underscore that in both China’s domestic and international governance, frontier AI safety is a secondary priority to AI development and economic growth.
Broader US-China tensions create structural obstacles to international cooperation. A particular stumbling block for US-China exchange came in March when BAAI was placed on the US Entity List, a designation generally justified by an organization’s ties to foreign military activity, and which prohibits organizations under US jurisdiction from trading with or transferring technology to the listed entity. Although the Entity List does not forbid non-commercial interaction, the exact boundaries on a technical topic like AI are not entirely clear, and a listing is sure to have a chilling effect.
CnAISDA's emergence demonstrates that shared technical concerns about AI risks can cross geopolitical boundaries. Moreover, the testing and evaluation work in CnAISDA’s constituent institutions, particularly Shanghai AI Lab, is a promising signal that China could be substantively working on measures to mitigate risks with global implications. This relative progress should not be taken as evidence of a high absolute level of concern or action: the AI safety ecosystem in China remains far less mature than its US counterpart in both technical and policy domains. Nonetheless, these developments illustrate how some coherence of governance can be achieved internationally via organic diffusion of best practices through locally grown institutions, even while overarching global frameworks remain elusive.
CnAISDA is a critical indicator of the influence of China’s frontier AI safety community. For the international community, CnAISDA represents an opportunity, but an uncertain one. It offers an authoritative formal channel for engaging China on AI safety. However, like the Chinese ecosystem more broadly, CnAISDA has so far taken little substantive action to address potentially global-scale risks. The truest test of CnAISDA’s enduring influence will be whether it can anchor a system-wide shift inside China toward not just speaking about frontier AI risks, but taking action to reduce them.
This article is largely based on a paper published by the same authors, available here.
See things differently? AI Frontiers welcomes expert insights, thoughtful critiques, and fresh perspectives. Send us your pitch.
Despite years of effort, mechanistic interpretability has failed to provide insight into AI behavior — the result of a flawed foundational assumption.
Congress is weighing a measure that would nullify thousands of state AI rules and bar new ones — upending federalism and halting the experiments that drive smarter policy.
