Overview & Analysis
The Measures for Labeling AI-Generated Synthetic Content establish China's dedicated regulatory framework for identifying and disclosing AI-generated material — covering text, images, audio, video, and virtual scenarios. The Measures require all qualifying network information service providers to apply both explicit labeling (visible or audible indicators that users can perceive directly) and implicit labeling (technical markers embedded in file metadata, including digital watermarks). The two labeling types work in tandem: explicit labels serve users who consume content, while implicit labels serve platforms and regulators that process or distribute it. The framework sits at the intersection of the Provisions on Deep Synthesis, the Algorithmic Recommendation Measures, and the Interim Measures for Generative AI Services — consolidating labeling obligations that were previously scattered across those instruments into a single, detailed operational standard.
This regulation is particularly important in the context of AI applications as generative AI technology is increasingly prevalent across news, advertising, entertainment, and education. To avoid misleading the public, maintain information transparency, and prevent the spread of false information, all service providers involved in generating content must comply with the labeling requirements. For AI projects, clearly labeling the source of generated content not only improves the platform's credibility but also strengthens user accountability and prevents the misuse of AI technologies. The Measures also impose obligations on content dissemination platforms — not just generators — making them a shared-responsibility framework across the AI content supply chain. Notably, users themselves are required to proactively declare and apply labels when publishing generated content, and deliberate deletion, alteration, forgery, or concealment of labels is explicitly prohibited.
Explicit Labeling (显式标识)
Visible or audible labels that users can directly perceive — required by service providers and declared by users.
- Text: textual prompts or symbols at start, end, or inline
- Audio: voice prompts or audio rhythm indicators
- Images: prominent indicators at appropriate positions
- Video: indicators in opening frame and around playback
- Virtual scenarios: indicators in initial frame
- Must be retained when downloaded, copied, or exported
Implicit Labeling (隐式标识)
Technical markers embedded in file metadata — not easily perceived by users but detectable by platforms and regulators.
- Mandatory: embedded in file metadata (Article 5)
- Encouraged: digital watermarks in content data
- Must include: content attributes, provider name/code, content ID
- Dissemination platforms check metadata to determine label obligation
- Platforms add dissemination chain info when forwarding labeled content
This policy applies to all scenarios involving AI-generated synthetic content — text, images, audio, video, and virtual scenarios. Both service providers that generate content and platforms that disseminate it carry labeling obligations. Users who publish generated content must also proactively declare and label it.
1. AI-Generated News, Advertising & Creative Content
When AI generates news reports, advertisements, or creative content, the generated content must be clearly labeled to avoid misleading the public. Explicit labels indicating AI origin are required at specified positions — the beginning, end, or appropriate locations within text; the opening frame and around playback for video. This labeling allows users to immediately identify the content's source and ensures informed consumption of AI-generated material.
2. AI-Generated Content on Social Media & Online Platforms
On social media, video platforms, and interactive online platforms, AI-generated content — such as comments, articles, and virtual character videos — must be clearly labeled. Dissemination platforms carry three-tier obligations: checking metadata for implicit labels, adding prominent indicators where metadata confirms generation, accepting user declarations where metadata is absent, and detecting explicit label traces or other generation signs to flag suspected generated content.
3. Virtual Scenarios and Metaverse Applications
In metaverse and virtual reality applications, AI-generated scenarios and interactive content must be labeled with explicit indicators at the initial frame and, optionally, during the course of the virtual service. These labels ensure users are aware whether the immersive content they are engaging with is AI-generated — particularly important as synthetic virtual environments become increasingly indistinguishable from human-created ones.
4. Content Generation Platforms Offering AI Creation Tools
Service providers offering AI creation tools and platforms must specify labeling methods and standards in user service agreements and prompt users to read and understand labeling requirements. Application distribution platforms must also require app providers to declare whether they offer AI-generated content services during listing reviews and verify labeling materials before apps go live. This creates a pre-market check that applies to the entire app ecosystem.
5. Cross-Border Content Sharing & International AI Services
When AI-generated content involves cross-border data sharing and distribution, service providers need to ensure that content is transparently labeled and that user privacy is protected through appropriate measures. Compliance with labeling requirements must extend across the content distribution chain, including international platforms and partners. Providers must also submit labeling-related materials when completing algorithm filings and security assessments, integrating labeling compliance into the broader AI regulatory process.
These Measures provide a standardized and transparent framework for labeling AI-generated content in China. For AI project managers, ensuring the legality and transparency of content and complying with Chinese regulations is key to reducing legal risks and enhancing platform trustworthiness.
Build Labeling into Content Pipelines by Content Type
The Measures specify different explicit labeling requirements for each content type: text, audio, images, video, and virtual scenarios each have distinct positioning and format rules. Do not apply a one-size-fits-all approach. Map your AI-generated content outputs by type and build type-specific labeling into the generation pipeline — not as a post-processing step but as an integral part of the output module. Ensure that downloaded, copied, or exported files retain the required explicit labels.
Implement Metadata Embedding and a Digital Watermarking Program
Implicit labeling in file metadata is mandatory; digital watermarks are explicitly encouraged. For multinationals, this means working with engineering teams to implement metadata tagging that includes content attributes, provider name or code, and content identifier — for every piece of AI-generated content, not just high-profile outputs. A digital watermarking program provides an additional layer that survives format conversions or screenshot circumvention, and supports the platform's ability to detect generation traces even when explicit labels have been removed.
Assign Labeling Responsibilities Clearly Across the Content Chain
The Measures impose obligations on generators, dissemination platforms, app stores, and users — not just on the service that generates the content. Multinationals operating across multiple roles (both generating and distributing AI content) must assign clear internal ownership for each obligation. Dissemination platforms must verify metadata and add their own chain information; app stores must verify labeling compliance at listing. Build these into cross-functional governance workflows, not just technical product requirements.
Establish Regular Labeling Compliance Reviews and Incident Procedures
As labeling standards evolve — the Measures reference other instruments such as the Deep Synthesis Provisions and Algorithmic Recommendation rules that may impose additional requirements — companies should establish periodic compliance reviews to track regulatory changes and update labeling implementations accordingly. Separately, Article 10's prohibition on malicious deletion, alteration, forgery, or concealment of labels means that companies should also have incident detection and response procedures for labeling tampering — including in user-generated content workflows where third parties may attempt to strip or forge labels.
Through early planning, rigorous implementation of content labeling, and compliance checks across the full AI content supply chain, companies can reduce legal risks while enhancing platform trustworthiness and promoting the healthy development of AI technologies in China. Labeling is not merely a disclosure obligation — it is the foundation of user trust in an era of increasingly convincing synthetic media.
Complete Regulatory Text
Article Index
- Articles 1–3 — Purpose, Scope & Definitions (Explicit and Implicit Labeling)
- Articles 4–5 — Explicit Labeling Requirements (by Content Type) & Implicit Labeling Requirements
- Article 6 — Labeling Obligations for Content Dissemination Platforms
- Articles 7–9 — App Distribution Platforms, User Agreements & Unlabeled Content Requests
- Articles 10–14 — User Obligations, Prohibitions, Compliance, Filing, Liability & Effective Date
Labeling of AI-generated synthetic content includes explicit labeling and implicit labeling.
"Explicit labeling" refers to labels added to generated content or interaction interfaces that are presented in forms such as text, audio, or graphics and can be clearly perceived by users.
"Implicit labeling" refers to labels embedded in the data of generated content files through technical means that are not easily perceived by users.
(1) For text: add textual prompts or common symbol indicators at the beginning, end, or appropriate positions within the content, or add prominent indicators in the interaction interface or around the text;
(2) For audio: add voice prompts or audio rhythm indicators at the beginning, end, or appropriate positions, or add prominent indicators in the interaction interface;
(3) For images: add prominent indicators at appropriate positions;
(4) For video: add prominent indicators at appropriate positions in the opening frame and around playback, and optionally at the end or in the middle;
(5) For virtual scenarios: add prominent indicators at appropriate positions in the initial frame and optionally during the service process;
(6) For other generative synthesis service scenarios: add prominent indicators based on application characteristics.
Where service providers offer download, copy, or export functions, they shall ensure that files contain compliant explicit labels.
Service providers are encouraged to include implicit labels such as digital watermarks in generated content.
"File metadata" refers to descriptive information embedded in the file header using specific encoding formats to record source, attributes, and usage information.
(1) Confirmed generated content: Verify whether file metadata contains implicit labels; where metadata clearly indicates generated content, add prominent indicators around published content to clearly inform the public that it is AI-generated;
(2) User-declared generated content: Where metadata lacks implicit labels but users declare the content as generated, add prominent indicators to inform the public that the content may be generated;
(3) Suspected generated content: Where metadata lacks implicit labels and users do not declare it, but the provider detects explicit labels or other signs of generation, identify it as suspected generated content and add prominent indicators;
(4) User declaration tools: Provide labeling tools and prompt users to declare whether content includes generated material.
In cases under items (1) to (3), metadata shall include content attributes, platform identifiers, and content identifiers to record dissemination chain information.
No organization or individual may maliciously delete, alter, forge, or conceal labels, provide tools for such actions, or use improper labeling methods to harm others' lawful rights.
人工智能生成合成内容标识办法
（2025年3月7日印发，自2025年9月1日起施行）
国信办通字〔2025〕2号 来源：中国网信网
条文索引
人工智能生成合成内容标识包括显式标识和隐式标识。
显式标识是指在生成合成内容或者交互场景界面中添加的，以文字、声音、图形等方式呈现并可以被用户明显感知到的标识。
隐式标识是指采取技术措施在生成合成内容文件数据中添加的，不易被用户明显感知到的标识。
（一）在文本的起始、末尾或者中间适当位置添加文字提示或者通用符号提示等标识，或者在交互场景界面、文字周边添加显著的提示标识；
（二）在音频的起始、末尾或者中间适当位置添加语音提示或者音频节奏提示等标识，或者在交互场景界面中添加显著的提示标识；
（三）在图片的适当位置添加显著的提示标识；
（四）在视频起始画面和视频播放周边的适当位置添加显著的提示标识，可以在视频末尾和中间适当位置添加显著的提示标识；
（五）呈现虚拟场景时，在起始画面的适当位置添加显著的提示标识，可以在虚拟场景持续服务过程中的适当位置添加显著的提示标识；
（六）其他生成合成服务场景根据自身应用特点添加显著的提示标识。
服务提供者提供生成合成内容下载、复制、导出等功能时，应当确保文件中含有满足要求的显式标识。
鼓励服务提供者在生成合成内容中添加数字水印等形式的隐式标识。
文件元数据是指按照特定编码格式嵌入到文件头部的描述性信息，用于记录文件来源、属性、用途等信息内容。
（一）核验文件元数据中是否含有隐式标识，文件元数据明确标明为生成合成内容的，采取适当方式在发布内容周边添加显著的提示标识，明确提醒公众该内容属于生成合成内容；
（二）文件元数据中未核验到隐式标识，但用户声明为生成合成内容的，采取适当方式在发布内容周边添加显著的提示标识，提醒公众该内容可能为生成合成内容；
（三）文件元数据中未核验到隐式标识，用户也未声明为生成合成内容，但提供网络信息内容传播服务的服务提供者检测到显式标识或者其他生成合成痕迹的，识别为疑似生成合成内容，采取适当方式在发布内容周边添加显著的提示标识，提醒公众该内容疑似生成合成内容；
（四）提供必要的标识功能，并提醒用户主动声明发布内容中是否包含生成合成内容。
有前款第一项至第三项情形的，应当在文件元数据中添加生成合成内容属性信息、传播平台名称或者编码、内容编号等传播要素信息。
任何组织和个人不得恶意删除、篡改、伪造、隐匿本办法规定的生成合成内容标识，不得为他人实施上述恶意行为提供工具或者服务，不得通过不正当标识手段损害他人合法权益。
