Measures for the Labeling of AI-Generated and Composed Content

Measures for the Labeling of AI-Generated and Composed Content.

人工智能生成合成内容标识办法

Promulgated by: CAC, MIIT, MPS, NRTA.
Document No.: Guo Xin Ban Tong Zi [2025] No. 2.
Issued March 7, 2025. Effective September 1, 2025.

Article 1. These Measures are enacted in accordance with the Cybersecurity Law of the People’s Republic of China, the Administrative Provisions on Algorithm Recommendation for Internet-based Information Services, the Administrative Provisions on Deep Synthesis of Internet-based Information Services, the Provisional Measures for the Administration of Generative Artificial Intelligence Services and other laws, administrative regulations and departmental rules for the purposes of promoting the healthy development of AI, regulating the labeling of AI-generated or composed content, protecting the legitimate rights and interests of citizens, legal persons and other organizations and safeguarding social and public interests. “ ” Article 2 These Measures shall apply to the labeling of AI-generated or composed content by Internet information service providers (hereinafter referred to as the “service providers”) in compliance with the Administrative Provisions on Algorithm Recommendation for Internet-based Information Services, the Administrative Provisions on Deep Synthesis of Internet-based Information Services and the Provisional Measures for the Administration of Generative Artificial Intelligence Services.

Article 3. The term “AI-generated or composed content” refers to the text, images, audio, video, virtual scenes and other information generated or composed by using AI technology. Labels for AI-generated or composed content include explicit label and implicit label. Explicit label refers to label added in the interface for generated or composed content or interactive scenes, which is presented in the form of text, sound or graphics and can be clearly perceived by users. Implicit label refers to label added to the file data of generated or composed content by taking technical measures, which is not easily perceived by users.

Article 4. Where the generation or synthesis services provided by a service provider fall under the circumstances stipulated in the first paragraph of Article 17 of the Administrative Provisions on Deep Synthesis of Internet-based Information Services, the service provider shall add explicit labels to the generated or composed content under the following requirements: (1) adding text prompts, general symbol prompts or other labels at the beginning, end or an appropriate position in the middle of the text, or adding visible prompt labels in the interactive scene interface or around texts;

(2) adding voice prompts, audio rhythm prompts or other labels at the beginning, end or an appropriate position in the middle of the audio, or adding visible prompt labels in the interactive scene interface;

(3) adding visible prompt labels at appropriate locations on the images;

(4) adding visible prompt labels at the beginning of the video and at appropriate locations around the video, or adding visible prompt labels at appropriate locations at the end and in the middle of the video;

(5) adding visible prompt labels at an appropriate location in the starting screen when the virtual scene is presented, and adding a visible prompt label at an appropriate location during the continuous service of the virtual scene; and

(6) adding visible prompt labels based on respective application characteristics for other generation or synthesis service scenarios. When service providers provide functions such as downloading, copying, and exporting generated or composed content, they shall ensure that the files contain explicit labels that meet the requirements.

Article 5. Service providers shall, in accordance with Article 16 of the Administrative Provisions on Deep Synthesis of Internet-based Information Services, add implicit labels to the file metadata of generated or composed content. Implicit labels shall include the attributes of the generated or composed content, the name or code of the service provider, content serial number and other production element information. Service providers are encouraged to add implicit labels in the form of digital watermarks or the like in the generated or composed content. File metadata refers to the descriptive information embedded in the file header in a specific coded format, which is used to record such information as the file’s source, attributes, and purpose.

Article 6. Service providers that provide online information content dissemination services shall take the following measures to regulate the dissemination of generated or composed content: (1) They shall verify whether there are implicit labels in the file metadata. If the file metadata explicitly indicates that it is generated or composed content, appropriate ways shall be taken to add visible prompt labels around the published content to explicitly remind the public that the content is generated or composed;

(2) If no implicit labels are verified in the file metadata, but the user declares that the content is generated or composed, appropriate ways shall be taken to add visible prompt labels around the published content to remind the public that the content may be generated or composed;

(3) If no implicit labels are verified in the file metadata, and the user has not declared that the content is generated or composed, but the service providers that provide online information content dissemination services detect explicit labels or other traces of AI-generation or synthesis, the content shall be identified as suspected generated or composed content, and appropriate ways shall be taken to add visible prompt labels around the published content to remind the public that the content is suspected of being generated or composed; and

(4) They shall provide necessary labeling functions and remind users to proactively declare whether the published content contains generated or composed content. Under the circumstances of Items (1) to (3) of the preceding paragraph, the attribute information of the generated or composed content, the name or code of the dissemination platform, the content serial number and other dissemination elements shall be added to the file metadata.

Article 7. When an application is put on shelves or made available online for review, an Internet application distribution platform shall require the Internet application service provider to state whether it provides AI-generation or synthesis services. Where the Internet application service provider provides such services, the Internet application distribution platform shall verify the materials related to the labels for its generated or composed content.

Article 8. Service providers shall specify in the user service agreements the methods, styles and other specifications for labeling generated or composed content and remind users to carefully read and understand the relevant labeling management requirements.

Article 9. Where a user requests a service provider to provide generated or composed content without any explicit labels, the service provider may, after specifying the user’s labeling obligations and use responsibilities in the user agreement, provide such content without any explicit labels and keep relevant logs such as information on the objects for not less than six months in accordance with the law.

Article 10. Users who use online information content dissemination services to publish generated or composed content shall proactively declare and use the labelling function provided by the service provider. No organization or individual may maliciously delete, tamper with, forge or conceal the labeling of generated or composed content as prescribed in these Measures or provide tools or services for others to commit the aforesaid malicious acts or damage the legitimate rights and interests of others by improper means of labeling.

Article 11. Service providers carrying out labeling activities shall also comply with the requirements of the relevant laws, administrative regulations, departmental rules and mandatory national standards.

Article 12. When performing the formalities for algorithm filing and security evaluation, service providers shall provide the relevant materials on the labeling of generated or composed content in accordance with these Measures, strengthen the sharing of labeling information and provide support and assistance for preventing and cracking down on relevant illegal and criminal activities.

Article 13. Whoever violates the provisions of these Measures shall be punished by the relevant competent authorities of cyberspace, telecommunications, public security, and radio and television ex officio and in accordance with the provisions of relevant laws, administrative regulations and departmental rules.

Article 14. These Measures shall come into force as of September 1, 2025.

§ COMMENTARY

Briefs on this law.

5 briefs reference this law.

§ 01 · AI-GOVERNANCE

China's AI-Companion Rule Takes Effect July 15 — A Clause-by-Clause Field Guide to What Actually Changed

China's Interim Measures for AI Anthropomorphic Interaction Services (人工智能拟人化互动服务管理暂行办法) — the world's first dedicated rule on 'companion'-style AI — take effect on 15 July 2026. This DCC brief synthesises three Chinese-language readings published in the days before the effective date: 数据合规肖大国's article-by-article practitioner walkthrough, 网安寻路人 (Hong Yanqing)'s multi-part work on how to scope anthropomorphic interaction (including his 'Sentiment Interaction Event / SIE' indicator system), and AI前沿信息笔记's read of the business-model logic the rule is really aimed at. Three throughlines: (1) what changed between the consultation draft and the final text — real fines were added, a 'continuity (持续性)' qualifier now narrows scope, the emergency-contact duty was widened beyond vulnerable groups, and the mandatory 'human takeover' of at-risk conversations was dropped; (2) the scope question the rule leaves under-specified — which services are 'continuous emotional interaction' at all — and the SIE-style indicator approach practitioners are reaching for to answer it; and (3) the paradigm shift the rule marks, from *content-safety* governance (AI as tool) to *relationship* governance (AI as social role), which finally gives regulators a handle on attention-economy and emotional-dependency business models. For overseas counsel shipping companion, emotional-AI or character-AI products into China: this is the operational checklist and the open-question list, two weeks out.

3 JUL 2026 ai-governance · companion-ai · anthropomorphic-ai
§ 02 · AI-GOVERNANCE

China's First AI-Ghostwritten 'Seeding Post' Case — a Duty of Care for Generative-AI Providers

China's first unfair-competition case over AI batch-ghostwritten 'seeding posts' (种草笔记 — the staged, first-person product-recommendation notes that drive discovery commerce on Xiaohongshu/RED). On appeal, the Hangzhou Intermediate People's Court ((2025) Zhe 01 Min Zhong No. 3998) held that the operators of an 'AI writing' tool ('AI写作鹅') that let users one-click-generate fake first-person Xiaohongshu notes — fabricating personal experiences and feelings — committed unfair competition under Article 2 (the general clause) of the Anti-Unfair Competition Law. The court built an explicit four-factor duty-of-care test for generative-AI providers (is it generative AI; does it target a specific scenario/another's product as its 'application layer'; is it directional and inducing; is it a paid, for-profit service), citing Articles 4(3), 5(1) and 22 of the Generative AI Services Interim Measures. Because the tool was named after Xiaohongshu, marketed to mass-produce on-brand 'seeding' copy, charged a membership fee, and shipped with no notice or reminder against the foreseeable misuse, the providers were at fault. The appeal court affirmed liability but cut damages from RMB 200,000 to RMB 100,000 on an 'inclusive and prudent' (包容审慎) view of AI, and reversed joint liability for the third defendant that merely hosted the download. DCC OCR'd the full judgment from the source images; this is our case brief for overseas counsel.

Updated · 8 JUN 2026 ai-governance · generative-ai · unfair-competition
§ 03 · AI-GOVERNANCE

China's First 'AI Hallucination' Tort Judgment — GenAI Is a Service, Not a Product, and the Chatbot's '¥100,000 Promise' Binds No One

The Hangzhou Internet Court has decided China's first 'AI hallucination' (AI幻觉) tort case — written into the Supreme People's Court's 2026 work report to the NPC. A user asking a chatbot about college applications was told, across seven rounds, that a non-existent campus existed; when finally shown the official website, the model 'apologised' and 'promised' to pay ¥100,000, even generating a fake lawsuit template telling him to sue. He did. The court dismissed every claim and, in doing so, laid down the first judicial articulation of China's generative-AI liability framework: (1) an AI model is not a civil subject, so its 'promise' is no declaration of intent — and is not attributable to the provider either; (2) generative AI is a service, not a product, so fault liability under Civil Code Article 1165 applies, not product liability's no-fault rule under Article 1202; (3) there is no result-based duty to guarantee accuracy for ordinary inaccurate output — only a process duty of care (conspicuous AI-content labelling plus industry-standard accuracy measures), which the provider had discharged; and (4) no proven damage, no causation. For any company deploying GenAI to the Chinese public, this is the operating liability surface and the evidentiary playbook.

4 JUN 2026 ai-governance · genai · ai-hallucination
§ 04 · AI-GOVERNANCE

Prompt Stacks and Prompt Governance — Why System-Level Prompts Are Emerging as a Regulatory Lever (and Where They Fall Short)

A Chinese AI-law reading of Neumann, Sargeant and Singh's FAccT 2026 paper Prompt Governance? — and what it means for how China, the EU, and the US treat 'system prompts' as a regulatory object. Li Wenlong (科技利维坦) walks through the four-layer 'prompt stack' (system instructions → system guidelines → developer instructions → user prompts), five properties practitioners need to understand (layered, hidden, natural-language, malleable, loosely coupled to behaviour), and the comparative regulatory landscape: the EU GPAI Code of Practice requires signatories to disclose system prompts to regulators in model reports; the Trump EO 14319 / OMB M-26-04 stops at model / system / data cards and leaves system-prompt disclosure voluntary; the UK's AI Cybersecurity Code says effectively nothing. China's current GenAI safety regime (TC260-003 plus the GenAI Interim Measures) is output-evaluation-based — filing and pre-launch scoring, with no architectural hook into system prompts. Li predicts a Brussels Effect: system-prompt disclosure to regulators will become a global compliance baseline, analogous to the DPIA in data law. For overseas counsel: this is what is coming, what to start archiving now, and why 'what you write' in a system prompt is not 'what the model executes.'

1 JUN 2026 ai-governance · system-prompts · prompt-stack
§ 05 · AI-AGENTS

Mapping the AI Agent Risk Surface — A Ten-Category Taxonomy Under China's New 智能体新规

China's Cyberspace Administration jointly issued the Implementation Opinions on Standardized Application and Innovation Development of AI Agents (the '智能体新规' or 'Agent Rules') on May 8, 2026 — the first dedicated regulatory document on AI agents anywhere in the world. This DCC brief works through the ten-category risk taxonomy that practitioners are now using to map the agent attack surface: goal hijacking, tool misuse, identity/permission abuse, supply-chain compromise, unintended code execution, memory and context poisoning, inter-agent communication insecurity, cascade failures, human-machine trust exploitation, and rogue agents. With the agent risk mapped, the brief works the legal-liability vector: how each risk maps to administrative, civil, and criminal exposure under existing PIPL, CSL, Anti-Unfair Competition, and trade-secret regimes. Closes with the Guangzhou Internet Court's recent dual-authorization ruling against an open-source agent that bypassed a chat platform's risk controls — the first Chinese case to articulate the dual-authorization principle for AI agents accessing third-party platforms.

28 MAY 2026 ai-agents · ai-governance · genai

Measures for the Labeling of AI-Generated and Composed Content.

China's AI-Companion Rule Takes Effect July 15 — A Clause-by-Clause Field Guide to What Actually Changed

China's First AI-Ghostwritten 'Seeding Post' Case — a Duty of Care for Generative-AI Providers

China's First 'AI Hallucination' Tort Judgment — GenAI Is a Service, Not a Product, and the Chatbot's '¥100,000 Promise' Binds No One

Prompt Stacks and Prompt Governance — Why System-Level Prompts Are Emerging as a Regulatory Lever (and Where They Fall Short)

Mapping the AI Agent Risk Surface — A Ten-Category Taxonomy Under China's New 智能体新规