SafetyRisk, alignment & guardrails
China's AI safety rules focus on what models say, and only 31 of 857 Chinese model releases came with published safety evaluations
China strictly enforces rules on what AI models say, while rules on frontier risk get much less enforcement. Regulators have started addressing agent and self-improvement risks, but only 3.6% of 857 releases from nine leading developers came with published safety evaluations.
From the Financial Times · AI article of November 16, 2026 and cited sources: Enforcement is concentrated at the application layer. [2] [7]
SOURCES[2] SemiAnalysis: Beijing Will Not Pace the Frontier newsletter.semianalysis.com[7] China Media Project: How China Sees AI Safety chinamediaproject.org
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Financial Times · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 16, 2026 article published
- Covered
- Nov 17 edition
- Evidence
- 10 sources across 9 sites
- 1What the west gets wrong about AI safety in Chinaft.com
- 2SemiAnalysis: Beijing Will Not Pace the Frontiernewsletter.semianalysis.com
- 3Concordia AI: State of AI Safety in China (2026)aisafetychina.substack.com
- 4AI Weekly: NYT on Chinese reading of Western AI warningsaiweekly.co
- 5Foreign Policy: China's AI Rollout Has No Off Switchforeignpolicy.com
- 6Jasmine Li: AI Safety in China, A Primerjasminexli.substack.com
Reach: not measured yet
Enforcement is concentrated at the application layer. Rules govern model outputs and behaviour, and the CAC's 2025 standards put violations of 'core socialist values' first among 31 listed risks. There are no compute thresholds and no duties to assess frontier risk before release. [2] [7]
Policy on frontier and agent risk is moving. In May 2026, the CAC and two other agencies issued guidance on agentic AI. In September, a CAC-affiliated standards body released Framework 3.0, which warns about recursive self-improvement. Chinese frontier-safety research output has also risen sharply. Technical rules and enforcement for the highest-risk tiers are still at an early stage. [3] [2]
Company-level transparency lags behind policy. Only 9 of 857 releases had evaluation results at launch. Concordia's count of 5 of 10 leading developers reporting evaluations measures developers rather than individual releases. Once open weights are out, as with Zhipu's GLM-5.3 after a two-week hold, they cannot be recalled. [2] [3] [5]
Western framing shapes how China receives safety messages. The FT argues that censorship is central to China's approach but that state security is not its only priority. Reporting indicates Chinese officials read Western warnings about existential risk as a ploy to slow Chinese labs, and hawkish messaging can undercut safety advocates inside China. [1] [4] [10]
The evidence shows strict content rules and sparse published evaluations.
Read the full assessment
For teams adopting Chinese open-weight models, this means running their own safety, cyber-capability and content-control tests instead of assuming pre-release checks took place.
Executive brief
The safety rules China enforces most strictly cover what AI models say. Rules on catastrophic or frontier risk get much less enforcement. A SemiAnalysis census found published safety evaluations for only 31 of 857 releases (3.6%) from nine leading Chinese developers between 2021 and September 2026. The FT story, What the west gets wrong about AI safety in China, argues that censorship is central to China's approach but that state security is not its only priority. Only a feed excerpt of the FT article is available, so its detailed evidence can't be assessed here. Other reporting shows real movement on agent and frontier-risk policy, while company-level transparency still lags.
What changed and event timeline
Regulators target agentic AI
The Cyberspace Administration of China (CAC) and two other agencies issued dedicated guidance on agentic AI. A full chapter covers safety and proposes risk-based governance. The trigger was the spread of the open-source OpenClaw agent.
Concordia publishes its annual census
The fourth edition reports that monthly Chinese frontier-safety papers rose about 60%. Agent safety reached 27% of new papers in Q1 2026.
Framework 3.0 released
TC260, a standards body under the CAC, released AI Safety Governance Framework 3.0. It warns of self-accelerating autonomous learning and recursive self-improvement.
Western warnings read as strategy
The New York Times reported that Chinese officials and media treat Western existential-risk warnings as culturally alien or as a ploy to slow Chinese labs.
Disclosure gap quantified
SemiAnalysis matched 857 model releases against published safety evaluations. Only 9 releases (1.1%) had results available at launch.
FT reframes the debate
The FT argues that censorship is central to China's approach but that state security is not Beijing's only AI-safety priority.
Capabilities and access
This is a governance story, not a model launch.
Read the full section
This is a governance story, not a model launch. Models named in the coverage:
- Frontier models: Kimi K3, Qwen 3.8 Max and GLM-5.3, described as "a few months behind" leading US models (Foreign Policy).
- GLM-5.3 release: Zhipu held back the weights for two weeks over concerns about the model's cyber capabilities, then released them. The model passed 1.4M Hugging Face downloads, and developers built roughly 80 variants within a month (same source).
- Censorship reach: DeepSeek's censorship also applies to global users of its product (Jasmine Li primer).
Technical analysis for researchers and developers
- China's rules are strict at the application layer: what a model may say or do.
- The CAC's March 2025 standards list 31 risks, with "core socialist values" violations at the top.
- AI Safety Benchmark 2.0 added categories for deception, loss of control and agentic risk (Concordia).
Read the full section
- Two different layers. China's rules are strict at the application layer: what a model may say or do. They have no compute thresholds or pre-release frontier-risk duties (SemiAnalysis).
- Ideology inside the safety taxonomy. The CAC's March 2025 standards list 31 risks, with "core socialist values" violations at the top. Compliance runs through training-data filtering and output labeling (China Media Project).
- Benchmarks. AI Safety Benchmark 2.0 added categories for deception, loss of control and agentic risk (Concordia).
- Reproducibility. SemiAnalysis's three datasets (a release census, 65 leadership statements, and 102 coded expert documents) are described in its analysis. Their availability for independent replication is not documented.
Claims and evidence
- Frontier-risk language is growing in policy texts, but technical specifications and enforcement for the extreme-risk tiers are still early (Concordia).
- Technical scientists prioritise frontier risk in 86% of their texts, while legal scholars prioritise development in 91%.
- Concordia counts 5 of 10 leading developers reporting evaluations with releases. SemiAnalysis counts 3.6% of releases.
Read the full section
- Frontier-risk language is growing in policy texts, but technical specifications and enforcement for the extreme-risk tiers are still early (Concordia).
- Expert opinion splits by profession. Technical scientists prioritise frontier risk in 86% of their texts, while legal scholars prioritise development in 91%. Top leadership decides (SemiAnalysis).
- Transparency figures conflict. Concordia counts 5 of 10 leading developers reporting evaluations with releases. SemiAnalysis counts 3.6% of releases. The figures measure different units (developers vs. individual releases), so they are not directly comparable.
- The FT's thesis cannot be checked against its full evidence, which is unavailable.
Context and prior work
- The China Media Project (July 2025) warned against assuming Chinese "AI safety" matches Western usage, citing chatbots that block questions about Uyghur cultural preservation.
- According to an AI Weekly summary of reporting on the topic, Xi Jinping chaired a Politburo study session on AI, and the national emergency response plan lists AI safety alongside pandemics (AI Weekly).
- An opposing view. Kristy Loke in The Wire China argues that Western analysis is missing China's catch-up on safety.
Read the full section
- Earlier warnings. The China Media Project (July 2025) warned against assuming Chinese "AI safety" matches Western usage, citing chatbots that block questions about Uyghur cultural preservation.
- High-level attention. According to an AI Weekly summary of reporting on the topic, Xi Jinping chaired a Politburo study session on AI, and the national emergency response plan lists AI safety alongside pandemics (AI Weekly).
- An opposing view. Kristy Loke in The Wire China argues that Western analysis is missing China's catch-up on safety.
Limitations, safety and contested findings
- Censorship vs. safety is disputed. Critics treat Chinese "safety" mainly as political control (China Media Project).
- Open weights can't be recalled. Releasing open weights removes any "off switch," whatever the regulation says (Foreign Policy).
- Western advocacy can backfire. Hawkish Western messaging can make safety advocacy harder inside China (LessWrong).
Read the full section
- Censorship vs. safety is disputed. Critics treat Chinese "safety" mainly as political control (China Media Project). Concordia documents genuine work on agent, companion-AI and frontier risk (Concordia).
- Open weights can't be recalled. Releasing open weights removes any "off switch," whatever the regulation says (Foreign Policy).
- Western advocacy can backfire. Hawkish Western messaging can make safety advocacy harder inside China (LessWrong).
Business and practitioner implications
- Check content controls before deploying. Chinese open-weight models carry built-in content controls tuned to PRC rules.
- For most Chinese releases, plan to run your own safety and cyber-capability evaluations (SemiAnalysis).
- Agent builders selling into China should track the CAC agentic guidance and the industry commitments on consumer agents (February 2026) and cloud agents (April 2026) (Concordia).
Read the full section
- Check content controls before deploying. Chinese open-weight models carry built-in content controls tuned to PRC rules. Red-team them for your own use case before deploying them globally.
- Don't assume evaluations exist. For most Chinese releases, plan to run your own safety and cyber-capability evaluations (SemiAnalysis).
- Agent builders selling into China should track the CAC agentic guidance and the industry commitments on consumer agents (February 2026) and cloud agents (April 2026) (Concordia).
- Policy engagement works better when framed around shared risks, such as agents and loss of control, than around slowing competitors.
The source trail.
Sources (10)
What the west gets wrong about AI safety in China
Feed excerpt only.
ft.com