⚡ Quick Summary
This research report identifies Chinese "AISI counterparts" — government-linked Chinese institutions doing similar work to the US and UK AI Safety Institutes (AISIs). No publisher, author or publication date is printed in the extracted text. The paper defines "first-wave" AISIs as government-backed technical institutions focused on the safety of advanced AI systems, with three core functions: technical research (including safety evaluations), standard-setting and cooperation.
Using a systematic review of open sources, three search methods and inclusion criteria requiring at least one core AISI function plus a government connection, the report names five most promising counterparts in a summary table: CAICT (evaluations), Shanghai AI Lab (technical research and evaluations, international cooperation), TC260 (standards), the Institute for AI International Governance (international cooperation) and BAAI (international cooperation). It also covers Peng Cheng Lab, TC28/SC42, CESA, the Cyberspace Administration of China and nascent bodies such as the Beijing Institute of AI Safety and Governance.
The report maps each institution's work — evaluation platforms such as Fangsheng and OpenCompass, benchmarks including the AI Safety Benchmark and SALAD-Bench, and standards such as TC260's AI Safety Governance Framework — and concludes that identifying promising counterparts does not amount to blanket approval of cooperation.
🧩 What’s Covered
- Executive summary and counterpart table: Table 1 lists the five most promising Chinese AISI counterparts — CAICT, Shanghai AI Lab, TC260, the Institute for AI International Governance and BAAI — against core AISI functions (technical research and evaluations, standards, international cooperation) and recommended engagement topics.
- Comprehensive institutional summary: an alphabetical, structure-grouped overview of every institution identified, from state-backed research institutions to standardization groups, the AI regulator and nascent bodies.
- Introduction and definitions: the US and UK AISIs, the International Network of AISIs and China's absence from it; the report's stated purpose of informing decisions about whom to engage and what to put on the agenda; and translation choices such as 安全 (ānquán) rendered as "safety/security".
- Method and inclusion criteria: three search methods (English-language reporting, Mandarin keyword searches on Google and Sogou, and Chinese government documents), the requirement for at least one core AISI function plus a government connection, and exclusions of academic and commercial groups and of funders.
- State-backed research institutions: BAAI's FlagEval platform and its "Safety and values" categories, the 2021 Technical Countermeasures for Security Risks of Artificial General Intelligence paper, standards contributions and convening roles (BAAI conferences, IDAIS, Beijing AI Principles); Peng Cheng Lab's cyber-range and military links; Shanghai AI Lab's evaluations (OpenCompass, MM-SafetyBench, From GPT-4 to Gemini and Beyond, FLAMES, SALAD-Bench, PsySafe) and its Safety Evaluations Working Group.
- CAICT, AIIA and AICTAE: the Large Model Governance Blue Paper and White Paper on Global Digital Governance, the Fangsheng platform with its AI Safety Benchmark and Responsibility/Safety Scores, the "Deep Alignment" project, AGI and AI-agent testing, and standards and working groups.
- I-AIIG and CISS: policy research, the annual International Forum on AI Cooperation and Governance, and CISS's track II dialogue with Brookings.
- Standardization groups, the regulator and emerging bodies: TC260's AI Safety Governance Framework and generative AI testing guidance, TC28/SC42's adoption of ISO/IEC 42001:2023, CESA's risk management capability assessment standard, CAC's pre-deployment review role, and new Beijing and Shanghai institutes and the Chinese AI Safety Network. An appendix reproduces FlagEval's "Safety and values" documentation.
💡 Why it matters?
For anyone deciding whether, with whom and about what to engage in China on AI safety, the report turns a scattered set of institutions, benchmarks and standards into a structured picture ordered by the functions AISIs themselves perform. It states that it can inform decisions about whom to engage and what to put on the agenda, and that it would be useful if a Chinese AISI were announced. It also flags counterparts it considers unsuitable, notably Peng Cheng Lab because of military links and the Cyberspace Administration of China because of its censorship role.
❓ What’s Missing
The extracted text prints no publisher, authors, publication date, version or document URL, so provenance cannot be checked from the document alone. The report excludes academic and commercial research groups and government funders by design, directing readers to Concordia AI databases instead, and covers institutions rather than the wider research literature. It acknowledges gaps: drafts of pre-trained model standards could not be accessed, the meaning of the AI Safety Benchmark's "AI consciousness" subcategories is unclear, and the degree of staff overlap between Shanghai AI Lab and SenseTime was not assessed. Emerging bodies are described only briefly.
👥 Best For
Policy analysts and AI governance researchers mapping China's AI safety ecosystem; officials in AI safety institutes, standards bodies and ministries planning international engagement; and risk or compliance teams that need to track Chinese evaluation platforms, benchmarks and AI standards as they develop.
📄 Source Details
Full title as printed: Chinese AISI Counterparts; the running footer on every page reads "CHINESE AISI COUNTERPARTS". No publisher, author, publication date, edition, series or reference number is printed in the extracted text, and no URL for the document itself appears — the only URLs are citations of other works. The input was the text extraction of all 73 PDF pages, but page 1 is blank, so cover details may not have been captured. The document is in English and reproduces Mandarin terms and source titles.
