China — Risk Classification
Relevant rules
Section titled “Relevant rules”| Rule | Legal tier | Relationship to risk classification |
|---|---|---|
| Generative AI Interim Measures (2023) | Departmental rule | Article 3 contains the “classified and graded supervision” principle |
| Deep Synthesis Provisions (2023) | Departmental rule | Sorts content by whether it “may cause public confusion” |
| Algorithmic Recommendation Provisions (2022) | Departmental rule | Distinguishes services “with public opinion attributes or capacity for social mobilisation” from ordinary ones |
| TC260-003-2024 | Technical specification | 31 risk categories; 90% safety pass rate |
| Science and Technology Ethics Review Measures (Trial) (2023) | Departmental rule | Article 25 sets a list that triggers expert review |
Three features of risk classification in China
Section titled “Three features of risk classification in China”1. There is a “classified and graded” principle, but no unified list
Section titled “1. There is a “classified and graded” principle, but no unified list”- Article 3 of the Generative AI Interim Measures states: 《生成式人工智能服务管理暂行办法》第三条 “对生成式人工智能服务实行包容审慎和分类分级监管” — i.e., “generative AI services shall be subject to inclusive, prudent, classified and graded supervision.”
- Article 21 of the Data Security Law sets three tiers for data: general / important / core.
- As of Apr 2026, no unified grading catalogue for “generative AI systems” has been issued.
Result: classification exists in name; in practice it is produced by stacking scenario-specific departmental rules.
2. Scenario-based ≠ risk-tiered
Section titled “2. Scenario-based ≠ risk-tiered”In practice China’s “tiers” are produced by peeling off scenarios and regulating each separately:
- Deep synthesis → one rulebook
- Generative AI → another
- Algorithmic recommendation → a third
- “Public-opinion attributes / social-mobilisation capacity” → heightened filing
This looks more like a list of scenarios than a risk-tier system. A unified cross-scenario intensity yardstick does not exist.
3. Filing (bei’an) as the de facto gatekeeper
Section titled “3. Filing (bei’an) as the de facto gatekeeper”- Models that meet certain capability or user-scale thresholds, or that carry public-opinion attributes, must complete algorithm filing.
- Filing review uses TC260-003 as its yardstick, which functions as a binary risk gate (pass / fail).
- Without a completed filing, a service cannot legally serve the public.
This is the sharpest structural difference from the EU’s four-tier approach: EU tiers produce differentiated obligations; China’s filing regime produces a yes-or-no answer.
The “list” in science and technology ethics review
Section titled “The “list” in science and technology ethics review”Article 25 of the Science and Technology Ethics Review Measures establishes an ex-ante list of research activities that must undergo expert re-review. The AI-related categories are:
- Algorithms with the capacity for social mobilisation or for shaping social consciousness.
- Automated decision-making in scenarios bearing on security or personal health.
- Human-machine integration systems with pronounced psychological or emotional influence.
This is the closest thing in China to the EU’s “high-risk list”, but it operates at the R&D stage, not at product launch.
Comparison with the EU and the US
Section titled “Comparison with the EU and the US”- Versus the EU: the EU has a single four-tier list; China does not. The EU looks at “systems”, China at “service scenarios”.
- Versus the US: China is more systematised at the regulatory-instrument level (multiple rules across sectors), but US state laws such as the Colorado AI Act actually carry a clearer definition of “high-risk AI system”.