China’s Generative AI Interim Measures
The Interim Measures for the Management of Generative AI Services (生成式人工智能服务管理暂行办法) took effect August 15, 2023. Seven agencies. 24 articles. Every GenAI provider serving the Chinese public must comply.
生成式人工智能服务管理暂行办法
Interim Measures for the Management of Generative AI Services
Who Must Comply
Article 2 defines the Measures’ jurisdictional reach. The regulation applies to any organization that uses generative AI technology to provide services generating text, images, audio, video, or other content to the public within mainland China.
Public-Facing Providers
Any provider using GenAI to generate and deliver text, images, audio, or video content to the public within mainland China. This includes chatbots, image generators, and AI writing tools.
Internal-Use Only
Organizations using GenAI for internal operations only are exempt. This covers enterprises, research institutions, and industry associations that do not offer AI services to the general public.
Overseas Providers
Overseas providers whose generative AI services are accessible to users within China are also subject to these Measures. Geographic accessibility, not corporate domicile, determines jurisdiction.
Prohibited Content Categories
Article 4 requires all generated content to uphold Core Socialist Values (核心社会主义价值观). This is a legal term defined in Chinese law, not a general aspiration. The following seven content categories are expressly prohibited. Tap each card for enforcement detail.
Subversion of State Power
Inciting overthrow of national sovereignty or the socialist system
Content that incites subversion of state power, overthrow of the socialist system, or actions aimed at splitting the state. This aligns with existing Criminal Law provisions and Cybersecurity Law content restrictions.
National Security
Endangering national security, interests, or the nation’s image
Content that endangers national security or national interests, or damages the nation’s image. The scope of “national security” is broadly defined under the National Security Law (2015).
Separatism
Inciting separatism or undermining national unity and social stability
Content that incites separatism, undermines national unity, or damages social stability. This category is enforced through both administrative and criminal channels.
Terrorism / Extremism
Promoting terrorism, extremism, ethnic hatred, or discrimination
Content promoting terrorism, extremism, ethnic hatred, ethnic discrimination, or religious extremism. Enforcement aligns with the Counter-Terrorism Law and regulations on religious affairs.
Violence / Obscenity
Violence, obscenity, or pornographic content generation
Content that is violent, obscene, or pornographic in nature. Providers must implement technical controls to prevent generation of such content and must log and report attempts.
False Information
Generating or disseminating false or harmful information
Content that generates or disseminates false information that may disrupt economic or social order. Under the Content Labeling Provisions (2025), AI-generated content must be labeled to prevent misidentification as human-authored.
Other Prohibited Content
Catch-all provision for content prohibited by other laws and regulations
A residual clause incorporating content prohibitions from all other applicable Chinese laws and administrative regulations, including the Advertising Law, Minor Protection Law, and sector-specific rules.
Draft vs. Final: Key Softening
The April 2023 draft drew immediate industry pushback from Baidu, Xiaomi, SenseTime, and government research institutes. The CAC described its revised approach as “inclusive and prudent” (包容审慎), applying “classified and graded” (分类分级) supervision. Three months later, the final text softened every major requirement.
Industry respondents included Baidu Xiaomi SenseTime and multiple government research institutes
Seven Issuing Agencies
The GenAI Interim Measures were jointly issued by seven agencies, the largest multi-agency AI regulation in China’s history. The CAC holds primary enforcement authority. Tap any card to see that agency’s specific role.
Cyberspace Administration of China. Lead drafter and primary enforcement body for all GenAI filing and compliance.
National Development and Reform Commission. Coordinates national AI industrial policy and investment planning.
Ministry of Education. Regulates GenAI use in educational settings and AI research at universities.
Ministry of Science and Technology. Oversees AI research ethics and innovation policy.
Ministry of Industry and Information Technology. Telecommunications licensing and AI hardware controls.
Ministry of Public Security. Law enforcement for cybersecurity violations and AI-related criminal activity.
National Radio and Television Administration. AI-generated content in broadcast media and online video.
Training Data Requirements
Article 7 establishes four core obligations for training data. The final version replaced the draft’s strict “true and accurate” standard with a more flexible quality-improvement mandate, but retained the source-contamination threshold and personal information consent requirements.
Providers must use high-quality training data and employ effective measures to improve data quality. The original draft demanded data be “true and accurate” (真实准确), which the industry argued was technically impossible for large language models. The final version adopted a process-based standard instead of an outcome-based one.
Art. 7(1)Training data must be legitimately sourced and must respect intellectual property rights. Providers bear responsibility for verifying that data collection does not infringe on copyright, patent, or other IP protections. This aligns with China’s Copyright Law and the 2024 Beijing Internet Court AI training data rulings.
Art. 7(2)Where training data contains personal information, providers must obtain explicit consent from the data subject. This requirement operates in conjunction with PIPL Articles 13-14, which define lawful bases for personal information processing. The PIPL consent standard is “informed, voluntary, and explicit.”
Art. 7(3) + PIPL Art. 13-14If a single data source contains more than 5% illegal or unhealthy content, the provider must not use that source for training. This creates a measurable, auditable threshold. Providers must implement automated content-filtering pipelines to detect and measure contamination levels before ingestion.
TC260 Implementation GuidanceMandatory Security Assessment
Article 17 requires providers of GenAI services with “public opinion properties or capacity for social mobilization” (舆论属性或社会动员能力) to complete a security assessment and algorithm filing with the CAC before launching their service to the public.
Algorithm Security Assessment
Evaluate the algorithm’s potential to generate prohibited content, produce biased outputs, or be exploited for harmful purposes.
Data Security Assessment
Assess training data sources, collection methods, storage practices, and compliance with PIPL and DSL requirements.
Content Security Assessment
Verify that output filtering, moderation systems, and content labeling meet the seven prohibited-category requirements under Article 4.
User-Facing Obligations
The Measures impose obligations not only on model providers but also on the user-facing layer. Providers must implement real-name registration, minor protection controls, and privacy safeguards for all users of their GenAI services.
Real-Name Registration
Users must register with verified identity: Chinese national ID or mobile phone number. Anonymous use of public-facing GenAI services is not permitted.
Art. 9Minor Protection
Providers must prevent addiction and over-dependence among minors. This requires age verification, usage time limits, and age-appropriate content filtering.
Art. 10Privacy and Data Rights
Minimize personal information collection. Protect user privacy. Grant users the right to review, correct, and delete their personal data held by the provider.
Art. 11Provider Compliance Checklist
The GenAI Interim Measures Art. 2 distinguishes between providers who develop models and providers who deploy models developed by others. Your obligations differ significantly:
- Model developers (those who train or substantially re-develop a foundation model) bear full compliance obligations: training data quality, keyword library, test banks, content filtering, and GenAI Filing with the CAC.
- API consumers/deployers (those using a pre-filed model via API) have lighter obligations: service-level content management, real-name verification, complaint handling, and GenAI Registration (not full filing) with their provincial CAC.
The checklist below covers the full set of obligations. API consumers can skip items marked with training data, keyword library, and test bank requirements if their upstream model already has an active CAC filing.
Track your compliance status against the Interim Measures’ core requirements. Click each item to mark it complete. For a downloadable version with additional detail, see the Templates page.
Continue Your Research
CAC Filing System
Three filing types, step-by-step registration, and the readiness assessment tool.
Explore →Data Triad: PIPL, DSL, CSL
How the three foundational data laws interact with GenAI training data obligations.
Explore →Content Rules and Labeling
The 2025 AI Content Labeling Provisions and mandatory watermarking requirements.
Explore →TC260 National Standards
83 published standards that define the technical safety thresholds for GenAI services.
Explore →Templates and Tools
Downloadable compliance checklists and filing decision trees for China’s AI regulations.
Explore →AI Governance Hub
Global frameworks, ISO 42001, NIST AI RMF, and the TJS 8-stage committee model.
Explore →Need Help with GenAI Filing?
TJS advisors help multinational teams prepare GenAI filing submissions, structure training data compliance, and build content safety testing programs.
Talk to a TJS Advisor →