Table of Contents
Quick Answer
China regulates AI through a vertical stack of Cyberspace Administration of China (CAC) rules in 2026: Algorithm Recommendation Provisions (2022), Deep Synthesis Provisions (2023), Interim Measures for Generative AI Services (August 2023), AI-Generated Content Labelling Measures (September 2025), and a draft national AI Law released by the Chinese Academy of Social Sciences for consultation.
- Every generative AI service for the Chinese public must file with the CAC
- All AI-generated content must carry explicit and implicit labels from 1 September 2025
- Training data must be lawful; content must uphold socialist values
What Is China's AI Regulatory Framework?
China's AI governance is issued jointly by the CAC, MIIT, Ministry of Public Security, and other regulators. The foundational instruments are the Cybersecurity Law (2017), Data Security Law (2021), and Personal Information Protection Law (PIPL, 2021). On top sit AI-specific rules administered primarily by the CAC.
The Interim Measures for the Management of Generative AI Services took effect 15 August 2023 (Order No. 15, CAC). The Administrative Measures for AI-Generated Synthetic Content Labelling (jointly issued by CAC, MIIT, MPS, and NRTA) took effect 1 September 2025. A comprehensive AI Law is on the State Council's 2025-2026 legislative agenda.
Key Details / Requirements
| Regulation | Year | Core Scope |
|---|---|---|
| Algorithm Recommendation Provisions | 2022 | Recommendation algorithms (news, e-commerce, short video) |
| Deep Synthesis Provisions | 2023 | Deepfakes, face-swap, voice cloning |
| Interim Measures for Generative AI Services | 2023 | Public-facing LLMs and image generators |
| Ethics Review Measures for S&T Activities | 2023 | Research ethics for high-risk AI |
| AI-Generated Content Labelling Measures | 2025 | Mandatory visible and embedded content labels |
| Draft AI Law | 2025-2026 | Comprehensive national statute (under review) |
Labelling Requirements (Effective 1 September 2025)
| Content Type | Explicit Label (human-readable) | Implicit Label (metadata) |
|---|---|---|
| AI-generated text | "AI-generated" text or icon visible to user | Required per GB/T 45438-2025 |
| AI-generated image | Visible watermark or caption | Required |
| AI-generated video | Persistent on-screen marker | Required |
| AI-generated voice | Spoken or text disclosure at start | Required |
Real-World Examples / Case Studies
Baidu's ERNIE Bot, Alibaba's Tongyi Qianwen, ByteDance's Doubao, and iFlytek's Spark — all completed mandatory CAC security assessments and algorithm filings before public launch in 2023-2024.
Tencent Hunyuan — Filed under the Deep Synthesis Provisions for its face-reenactment feature.
Meitu — Fined CNY 500,000 in 2024 for using user photos to train AI without explicit consent, under PIPL Article 66.
What This Means for Businesses
Any company offering generative AI to users in mainland China — domestic or foreign-invested — must:
- File the algorithm under the CAC Algorithm Registry
- Complete a CAC Security Assessment if the service has "public opinion properties or social mobilisation capabilities"
- Implement real-name user verification
- Label all AI-generated content per GB/T 45438-2025
- Maintain training data provenance records
- Refuse outputs that "endanger national security" or "contradict socialist core values"
Compliance Checklist
- Register the algorithm via the CAC's public filing portal
- Complete a Personal Information Protection Impact Assessment (PIPIA) per PIPL Article 55
- Implement explicit and implicit content labels before 1 September 2025
- Obtain separate consent for biometric training data
- Establish a mechanism for users to report illegal content within 24 hours
- Store all service logs for at least 6 months
Conclusion
China has one of the world's most comprehensive AI rule-stacks. Companies targeting Chinese users must treat algorithm filing, content labelling, and PIPL compliance as gating requirements.
Misar AI's China AI compliance guide is updated monthly — visit misar.blog/china-ai.