AI Policy Wiki
Dashboard

Concordia AI

medium confidence · updated 2026-07-25

Beijing-based AI safety and governance organization whose annual State of AI Safety in China surveys are the most widely cited source on Chinese frontier-lab safety disclosure. Its finding that only 3 of 13 surveyed Chinese labs published dedicated safety sections, and none disclosed CBRN or loss-of-control evaluations, is reproduced across US policy advocacy.

Concordia AI is a Beijing-based organization working on AI safety and governance, best known for its periodic State of AI Safety in China reports. It occupies an unusual position in the policy literature: a China-based organization whose survey findings are the standard citation in US and UK arguments that Chinese frontier labs disclose less safety information than their Western counterparts.

Activities

The organization's principal published output tracked here is the State of AI Safety in China report series, which surveys Chinese frontier AI developers on published safety practice. Its 2025 edition surveyed 13 Chinese frontier AI labs and found that nine had published technical model cards, only three had dedicated AI safety sections, and none disclosed evaluation results for CBRN or loss-of-control risks (Mind the US-China Safety Gap (OpenAI Global Affairs, July 2026)).

Concordia AI researchers also contribute to multi-institution technical governance work; the organization is among the 17 institutions represented in the authorship of Open Problems in Frontier AI Risk Management, alongside SaferAI, the AI Standards Lab, The Future Society, Pivotal Research, and academic governance centers.

Citation in policy argument

The 3-of-13 disclosure finding has become a recurring premise in US frontier-safety advocacy rather than a contested empirical claim. Anthropic cites it in 2028: Two Scenarios for Global AI Leadership (Anthropic) alongside a CAISI finding that DeepSeek's R1-0528 complied with 94% of overtly malicious requests under a common jailbreak against 8% for US reference models, and an independent assessment of Moonshot's Kimi K2.5 refusal rates on CBRN-related requests. OpenAI cites the same survey as the disclosure half of its "Mind the US-China Safety Gap" argument, pairing it with the company's own red-team comparison. See US-China AI Competition: Different Races, Different Metrics.

The finding is about published safety evaluation, not about internal practice, and the reports do not claim to observe unpublished evaluation work at the surveyed labs.

Relationships