⚡ Quick Summary
Published by Concordia AI, this Spring 2024 report updates the organisation's October 2023 survey of frontier AI safety in China. It defines "frontier AI" to include both "highly capable general-purpose foundation models" and narrow AI systems in dangerous domains, and states exclusions such as AI security topics (model weights, model data, development facilities) and lethal autonomous weapons.
The report surveys Chinese technical safety research, international governance, domestic governance, and lab and industry practices. It reports nearly 15 Chinese frontier AI safety papers per month over the previous six months against 6 per month for the preceding seven, identifies 11 key research groups, and finds alignment the most represented research direction while interpretability of frontier models is comparatively lacking.
Governance findings cover China's signing of the Bletchley Declaration, co-sponsorship of the first UNGA AI resolution, a joint statement on AI with France, and a new China–US AI dialogue whose first meeting is set for 14 May in Geneva. Domestically it examines two expert drafts of a national AI law, the TC260 generative AI security standard, NSFC value-alignment funding calls, and local AGI policies in six provinces and municipalities.
🧩 What’s Covered
- Introduction and scope: defines the report's subject as "frontier AI risks" — highly capable general-purpose foundation models plus narrow AI systems in dangerous domains — and lists exclusions, including AI security topics such as model weights, model data and development facilities, and lethal autonomous weapons.
- Technical safety research: sets out the dataset of frontier safety papers with substantial Chinese authorship from April 2023 to May 2024, the taxonomy adapted from Dan Hendrycks et al. (alignment, robustness, systemic safety, monitoring), the rise from 6 to nearly 15 papers per month, the 11 key research groups selected on anchor-authorship, and notable work on LLM unlearning, multi-agent alignment, backdoor and multimodal attacks, chemical and biological misuse, watermarking, and benchmarks such as SALAD-Bench, OpenEval and FLAMES.
- International governance: tracks China's signature of the Bletchley Declaration, co-sponsorship of the March 2024 UNGA AI resolution and its frontier-safety provisions in section 6, the China–Africa forum and its planned policy dialogue, the Sino-French joint statement, the China–US AI dialogue due to start on 14 May in Geneva, two WDTA standards on generative AI and LLM security testing, and the Track 1.5/2 dialogue landscape.
- Domestic governance: examines the 2024 Government Work Report's "AI+" initiative, the CASS-led and CUPL-led expert drafts of a national AI law and their provisions, the absence of new binding AI regulations, 117 registered generative AI products and over 900 deep synthesis algorithms, NSFC value-alignment funding calls, MSS and CAICT publications, the TC260 generative AI security standard with its 4,000-sample, 96% compliance test, and local AGI policies in six provinces and municipalities.
- Lab and industry practices: covers AIIA's Safety and Security Governance Committee and its AI safety benchmark, CSAC's expert committee and Chinese corpus, the safety disclosures published with InternLM2, ChatGLM and DeepSeek-V2, and Ant Group's claim that nearly 20% of its large-model technical team works on ethics.
💡 Why it matters?
For teams that must factor Chinese AI policy into risk assessments, the report maps who publishes frontier safety research in China, which ministries, standards bodies and alliances are active, and where official regulation ends and expert drafts, industry benchmarks and local pilot policies begin. It helps readers distinguish proposals from adopted rules. The document itself connects Chinese developments to international processes such as the Bletchley Declaration, the UNGA resolution and the Sino-French statement. It also records that only 2 of 8 AI-focused Track 1.5 and 2 dialogues since 2022 centred on frontier AI safety and governance, and notes that corporate safety claims such as Ant Group's cannot be independently verified.
❓ What’s Missing
The report states several limits of its own evidence. The paper dataset covers only frontier models, excluding safety-relevant research on smaller models, and the researcher-quality metrics are described as "incomplete and lagging". The TC260 generative AI security standard makes reference to long-term risks such as deception and self-replication but "has no concrete measures for these risks". Corporate AI ethics practices "remain largely a black box", and Ant Group's staffing claim is not verifiable. The two national AI law texts are unofficial expert drafts, and the survey stops at May 2024.
👥 Best For
Best for policy analysts, government-affairs and compliance teams tracking China's AI rulemaking and its international engagement, for AI safety researchers and funders mapping Chinese research groups and benchmarks, and for organisers of bilateral or multilateral AI safety dialogues who need a structured overview of Chinese activity through May 2024.
📄 Source Details
The State of AI Safety in China: Spring 2024 Report, dated 14 May 2024 and produced by Concordia AI, identified through internal references such as "our CEO, Brian Tse" and "Concordia AI collected a dataset". English, 114 pages, no series or reference number printed in the extracted pages. The input was a partial text extraction covering pages 1–80 of 114 (table of contents and sections 1–5); sections 6–9 were not available. No document URL is printed in the extracted pages.
