Interim Measures for the Administration of Generative Artificial Intelligence Services
English · 中文
生成式人工智能服务管理暂行办法 — CAC with six other departments, Order No. 15: adopted at the CAC's 12th executive meeting of 2023 on 23 May 2023, signed and promulgated 10 July 2023, in force 15 August 2023. Five chapters, twenty-four articles. China's first national-level rule addressed to generative AI services, and still the general statute that the later, narrower rules hang from.
In one line
Whoever offers a generative AI service to the public inside mainland China is a content producer under Chinese law: the training data must be lawful, the output must not contain listed categories of illegal content, unlawful output must be stopped at the generation stage and reported, services with public-opinion or social-mobilisation capacity need a security assessment plus an algorithm filing — and the measure itself sets no fine ceiling of its own.
Who and what it governs
- Scope (Article 2) — using generative AI technology to provide the public inside mainland China with services that generate text, images, audio, video and similar content. Where other rules govern generative AI used for publishing, film and television, or literary and artistic creation, those rules prevail.
- Excluded (Article 2, third paragraph) — industry associations, enterprises, education, science and research, public cultural institutions and other professional bodies that research or apply the technology without offering a generative AI service to the domestic public fall outside the measure.
- Providers include API sellers (Article 22) — a "provider" is any organisation or individual offering the service including by way of a programmable interface. A model exposed through an API is in scope even if the operator never sees an end user.
- Users are defined separately (Article 22) — organisations and individuals using the service to generate content; several duties run against them through the provider's contract.
Key provisions (the ones that change behaviour)
| Article | What it requires |
|---|---|
| 3 | The state pursues development and security in tandem, encouraging innovation while governing generative AI services with 包容审慎 (inclusive, prudent) and 分类分级 (classified and tiered) regulation. |
| 4 | Five obligations on providing and using the service: uphold socialist values and generate none of the listed prohibited content; take effective measures against discrimination on ethnicity, belief, country, region, gender, age, occupation or health in algorithm design, training-data selection, model generation and optimisation, and service provision; respect intellectual property and commercial ethics, and do not use algorithm, data or platform advantage to monopolise or compete unfairly; respect others' lawful rights — no harm to physical or mental health, no infringement of portrait, reputation, honour, privacy or personal-information rights; and, matched to the service type, raise transparency and the accuracy and reliability of generated content. |
| 5–6 | Encouragement, not obligation: applications across industries, cooperative work on data and risk, independent innovation in algorithms, frameworks, chips and supporting software, international exchange and participation in international rule-making, training-data and compute infrastructure, and orderly tiered opening of public data to widen high-quality public training corpora. |
| 7 | Pre-training and optimisation training must: use data and base models of lawful origin; not infringe intellectual property; obtain individual consent or another lawful basis where personal information is involved; take effective measures to raise training-data quality and its truthfulness, accuracy, objectivity and diversity; and meet the Cybersecurity Law, Data Security Law, Personal Information Protection Law and other regulator requirements. |
| 8 | Where data annotation is done in the course of R&D, the provider must set clear, specific and operable annotation rules, run quality assessment with sample verification of annotated content, and train and supervise the annotators. |
| 9 | The provider bears the responsibility of a network information content producer and its network-security duties; where personal information is involved it bears personal-information handler responsibility. A service agreement with each registered user sets out rights and obligations. |
| 10 | State who the service is for, in what settings, for what purposes, and publish it; guide users toward scientific, rational and lawful use; take effective measures to prevent minor users from over-dependency on or addiction to the service. |
| 11 | Protect users' input information and usage records: do not collect personal information that is not necessary, do not unlawfully retain input or records that identify the user, do not unlawfully provide them to others; and handle requests to access, copy, correct, supplement or delete personal information promptly. |
| 12 | Label generated images, video and similar content in accordance with the Provisions on the Administration of Deep Synthesis — the labelling duty arrives by cross-reference, not by a new standard here. |
| 13 | Provide a safe, stable and continuous service so users can rely on it. |
| 14 | On discovering illegal content: promptly stop generating and stop transmitting it, eliminate it, remediate through optimised model training, and report to the competent authorities. On discovering a user committing unlawful acts through the service: warn, restrict functions, suspend or terminate service to that user by law and by contract, keep records, and report. |
| 15, 18 | Build complaint and reporting mechanisms with convenient entry points, published process and response deadlines; users may complain to the competent authorities about a service that breaches the law or these measures. |
| 16 | CAC, development-and-reform, education, science and technology, industry-and-information-technology, public security, broadcasting and press-and-publication departments regulate within their remit; competent departments are directed to develop classified and tiered regulatory rules or guidance suited to the technology. |
| 17 | A service with public-opinion attributes or social-mobilisation capacity must carry out a security assessment under national rules and perform algorithm filing, amendment and cancellation under the Provisions on the Administration of Algorithmic Recommendations. |
| 19 | Providers must cooperate with supervision and, on request, explain the source, scale, type and annotation rules of training data and the mechanism of the algorithm, supplying necessary technical and data support. Institutions and personnel taking part in assessments must keep state secrets, trade secrets, personal privacy and personal information confidential. |
| 20 | Where generative AI services are supplied from outside mainland China into it and do not meet Chinese law or these measures, the national cyberspace administration notifies the relevant bodies to take technical and other necessary measures — the statutory hook for blocking at the infrastructure layer. |
| 21 | Penalties. Where the Cybersecurity Law, Data Security Law, Personal Information Protection Law, Science and Technology Progress Law and other laws provide, those provisions are applied; where they do not, the competent departments give a warning or circulated criticism and order correction within a time limit, and on refusal to correct or in serious cases order suspension of the relevant service. Public-security and criminal liability follow where conduct amounts to those offences. |
| 23 | Where laws or administrative regulations require an administrative licence to provide the service, obtain it; foreign investment in generative AI services must comply with the foreign-investment rules. |
What the penalty article does not do
Article 21 sets no independent fine amount. It routes punishment into the general laws — which is where the money actually is: the Personal Information Protection Law reaches RMB 50 million or 5% of prior-year turnover for serious violations (Article 66), and the amended Cybersecurity Law, in force 2026-01-01, reaches RMB 10 million with up to RMB 1 million on responsible individuals — on two separate triggers: Article 61 for network-security failures and Article 69 for prohibited information, both transcribed in the Cybersecurity Law card — and leaves the measure's own sanction to be the operational one: warning, circulated criticism, order to correct, order to suspend the service. This is the answer to the question left open in ../compare/01-ai-labeling.md: the generative-AI measures themselves carry no fine ceiling.
Relationship to other laws
- Named as its legal basis (Article 7(5), Article 21) — Cybersecurity Law, Data Security Law, Personal Information Protection Law, and the Science and Technology Progress Law.
- Article 12 imports the deep-synthesis labelling duty, and Article 17 imports the filing machinery of the Provisions on the Administration of Algorithmic Recommendations together with the 2018 Security Assessment Provisions whose Article 2 defines the trigger. The 2025 Measures for Labelling AI-Generated Synthetic Content later displaced the labelling detail; the Anthropomorphic Interaction Services Measures of 2026 add a companion-service layer on top.
- The Provisions on the Administration of Deep Synthesis (Order No. 12, effective 2023-01-10) remain in force alongside and are where the labelling duty of Article 12 actually comes from.
- Against the EU: the EU's comparable layer is the AI Act's general-purpose AI model regime, whose obligations start to apply 2025-08-02 — a model-side chapter (documentation, training-data summaries, copyright policy, systemic-risk duties) rather than a service-side licence. China regulates the service offered to the public and reaches the model through it; the AI Act regulates the model and reaches the service through its deployer rules. See
../eu/ai-act.md.
Why this matters for the comparison
This measure is why "China has no AI law" is misleading: it is not a statute, but from 15 August 2023 it has been the general obligation set for anyone publishing a generative service in China, and every narrower rule since (labelling, ethical review, anthropomorphic services) has been written to plug into it rather than to replace it. The EU's general statute came earlier in calendar terms but applies its model duties later than China's service duties took effect.
Official sources
- Full text, CAC (Order No. 15, promulgated 2023-07-10, posted 2023-07-13): https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm
- Full text, State Council policy document library (七部门令第 15 号, 成文日期 2023-07-10): https://www.gov.cn/zhengce/zhengceku/202307/content_6891752.htm
- Joint announcement by the seven departments (CAC news item): https://www.cac.gov.cn/2023-07/13/c_1690898326795531.htm
Source quotes
Every sentence below is copied from the official text and pinned to the snapshot in snapshots/ that this card was written against. node scripts/check-quotes.mjs re-checks them offline, so a quote that stops matching the source fails a check instead of quietly becoming the record. The prose above is a rendering; these are the anchors.
提供者违反本办法规定的，由有关主管部门依照《中华人民共和国网络安全法》、《中华人民共和国数据安全法》、《中华人民共和国个人信息保护法》、《中华人民共和国科学技术进步法》等法律、行政法规的规定予以处罚— Article 21 names four statutes and stops (snapshot:snapshots/cn/generative-ai-services-measures.txt)法律、行政法规没有规定的，由有关主管部门依据职责予以警告、通报批评，责令限期改正；拒不改正或者情节严重的，责令暂停提供相关服务— the fallback ladder — no fine amount anywhere (snapshot:snapshots/cn/generative-ai-services-measures.txt)第十七条提供具有舆论属性或者社会动员能力的生成式人工智能服务的，应当按照国家有关规定开展安全评估，并按照《互联网信息服务算法推荐管理规定》履行算法备案和变更、注销备案手续— Article 17 — assessment and filing (snapshot:snapshots/cn/generative-ai-services-measures.txt)提供者应当按照《互联网信息服务深度合成管理规定》对图片、视频等生成内容进行标识— Article 12 imports the labelling duty (snapshot:snapshots/cn/generative-ai-services-measures.txt)
Unverified / TODO
- The English rendering of every article name and term is this repository's own translation — no official English text was found. ⚠ [gap:no_official_translation]
- The security assessment required by Article 17 runs through the Provisions on the Security Assessment of Internet Information Services with Public-Opinion Attributes or Social-Mobilisation Capacity — provisions, not measures, and 2018 rather than 2022, which is what this card used to say. Identified, transcribed and carded; text checked against the CAC full text → card.
- Read the drafters' Q&A for this measure and record the official reading of "public-opinion attributes or social-mobilisation capacity" — it decides who files. ⚠ [gap:interpretation_missing]
- Check whether the practice guides and the mandatory generative-AI security standard issued since 2023 change what Article 7 and Article 17 require in practice — neither is carded here and no number has been verified for them. ⚠ [gap:not_located]
- Close the ⚠ in
compare/01-ai-labeling.mdandcompare/05-liability-and-pending-law.mdabout generative-AI penalties using Article 21 above — done in the same pass; both files now state that Art 21 enumerates four general laws and supplies its own ladder, while Art 22 of the deep-synthesis provisions refers generally with no fallback.