China's mandatory national standards on AI generated content labeling
morning meaningful guidance in terms of compliance
Yesterday, I introduced the Labeling Measures for AI-generated and Synthetic Content jointly issued by the Cyberspace Administration of China (CAC) and several other ministries. I have been closely following this labeling rule since its draft was released in September last year.
A mandatory national standard, Cybersecurity Technology—Methods for Labeling AI-Generated and Synthetic Content, was introduced almost simultaneously with the issuance of the Measures.
China's national standards system includes mandatory standards (强制性标准) and recommended standards (推荐性标准), both issued by the Standardization Administration of China (SAC). Mandatory standards usually begin with "GB," while recommended standards use "GB/T." "GB" stands for Guo Biao(国标), the pinyin abbreviation for national standard (Guo Jia Biao Zhun, 国家标准). Both represent national-level technical norms and standards applicable nationwide.
Mandatory national standards carry an authority almost equivalent to law and regulations. In fact, their practical significance for frontline compliance personnel in tech companies and other relevant entities can exceed that of legal provisions, as they provide detailed guidance on various compliance-related technical issues. By offering concrete and intuitive instructions, they demonstrate how AI content labeling is implemented in practice in China.
Recognizing the industry's frustration over ambiguous regulations that lack practical guidance, Chinese regulators chose to issue both a mandatory national standard and another technical document—Cybersecurity Standard Practice Guidelines: AI-Generated Synthetic Content Identification—Encoding Rules for Service Providers—simultaneously with the regulation (I will introduce this in my next article).
This marks a significant advancement in China’s AI regulatory approach, reflecting a shift from broad and generalized to more refined and precise regulation. It also offers a more concrete "Chinese model" for global AI governance—one that extends beyond vague policy language and diplomatic rhetoric. This approach demonstrates the continuous learning and exploration of China’s AI regulators in the course of prudent regulatory practice, and it is reasonable to expect that it will be carried forward in future AI governance efforts in China.
I have translated the full text of this mandatory national standard and hope it will be helpful to everyone.
Cybersecurity Technology: Labeling Method for Content Generated by Artificial Intelligence
ICS35. 030 CCSL80
People's Republic of China National Standard GB45438—2025
2025-02-28 Issued
2025-09-01 Implementation
Issued by National Market Supervision Administration National Standardization Management Committee
Foreword
This document is drafted in accordance with the rules of the "Standardization Work Guide Part 1: Structure and Drafting Rules for Standardization Documents" (GB/T1. 1—2020).
Note that some of the contents of this document may involve patents. The issuing body of this document does not bear the responsibility for identifying patents.
This document is proposed and managed by the Office of the Central Cybersecurity and Informatization Commission.
Introduction
To regulate the labeling activities of content generated by artificial intelligence by providers of generation and synthesis services and content dissemination services, according to the "Labeling Measures for Content Generated by Artificial Intelligence Synthesis" and other relevant laws and regulations, this standard for explicit labeling and implicit labeling methods is formulated to support the effective implementation of relevant laws and regulations.
Cybersecurity Technology Labeling Method for Content Generated by Artificial Intelligence
1. Scope
This document specifies the labeling methods for content generated by artificial intelligence.
This document applies to providers of generation and synthesis services and content dissemination services conducting labeling activities for content generated by artificial intelligence.
2. Normative References
The contents of the following documents, through normative references in this document, constitute indispensable terms of this document. For the referenced documents with dates, only the versions corresponding to the dates apply to this document; for those without dates, the latest versions (including all amendments) apply to this document.
GB18030—2022 Information Technology Chinese Character Coding Set
3. Terms and Definitions
The following terms and definitions apply to this document.
3.1 Content Generated by Artificial Intelligence
Content generated, synthesized by artificial intelligence technology, such as text, images, audio, video, virtual scenes, etc.
3.2 Label of Content Generated by Artificial Intelligence
Label indicating that a certain content is generated by artificial intelligence (3.1).
Note: The label may include information such as providers of generation and synthesis services and content dissemination services.
3.3 Explicit Label
Label added in content generated by artificial intelligence or in the interaction scenario interface, presented in text, sound, graphics, etc., and perceptible to users.
3.4 Implicit Label
Label added technically in the file data of content generated by artificial intelligence, not easily perceptible to users.
3.5 File Metadata
Descriptive data embedded in a file in a specific encoding format, recording information such as file source, attributes, purpose, and copyright.
3.6 Implicit Label in File Metadata
Implicit label added in the file metadata of content generated by artificial intelligence.
3.7 Provider of Artificial Intelligence Content Generation Services Provider of generation and synthesis services
An organization or individual providing generation and synthesis services (including through programmable interfaces) to the public using artificial intelligence technology.
3.8 Internet Information Content Propagation Service Provider Content dissemination service provider
A network information service provider offering network information content dissemination services.
4. Overview
Labeling of content generated by artificial intelligence includes explicit labeling and implicit labeling. Explicit labeling refers to labels added in content generated by artificial intelligence or in interaction scenario interfaces, presented in text, sound, graphics, etc., and perceptible to users, mainly to indicate to the public that the content is generated by artificial intelligence; implicit labeling refers to labels added technically in the file data of content generated by artificial intelligence, not easily perceptible to users, mainly to record information related to the generation and synthesis of content. Labeling methods are shown in Appendix A.
Explicit labeling can be divided into content explicit labeling and interaction scenario interface explicit labeling according to the labeled object. Content explicit labeling can be further divided into text, image, audio, video, virtual scene, etc. explicit labeling, and interaction scenario interface explicit labeling can be further divided into explicit labeling near the content and explicit labeling in an appropriate position of the interaction scenario interface. Typical application scenarios of explicit labeling are shown in Appendix B.
Implicit labeling can be divided into implicit labeling in file metadata, implicit labeling in content (such as digital watermark added in generated and synthesized content), etc., according to the labeling position.
5. Explicit Labeling
5.1 Text Content Explicit Labeling
Methods for explicit labeling of text content are as follows.
a) Explicit labeling of text content should be in the form of text or superscript.
b) Text-form explicit labeling of text content should simultaneously include the following elements:
1) Artificial intelligence element: including "artificial intelligence" or "AI", indicating the use of artificial intelligence technology;
Note: "AI" is the abbreviation of Artificial Intelligence.
2) Generation and synthesis element: including "generation" and/or "synthesis", indicating the content creation method as generation and/or synthesis.
c) Superscript-form explicit labeling of text content should include "AI".
d) Explicit labeling of text content should be located in one or more of the following positions:
1) At the beginning of the text;
2) At the end of the text;
3) At an appropriate position in the middle of the text.
e) The font and color of the explicit labeling of text content should be clear and distinguishable.
Examples of explicit labeling of text content are shown in Appendix C, C.1.
5.2 Image Content Explicit Labeling
Methods for explicit labeling of image content are as follows.
a) Explicit labeling of image content should be in the form of text prompts.
b) Explicit labeling of image content should simultaneously include the following elements:
1) Artificial intelligence element: including "artificial intelligence" or "AI", indicating the use of artificial intelligence technology;
2) Generation and synthesis element: including "generation" and/or "synthesis", indicating the content creation method as generation and/or synthesis.
c) Explicit labeling of image content should be located at the edge or corner of the image.
d) The font of the explicit labeling of image content should be clear and distinguishable.
e) The height of the text in the explicit labeling of image content should not be less than 5% of the length of the shortest side of the image.
Note: For non-rectangular images, the shortest side refers to the shortest side of the rectangle that can completely contain the image. Examples of explicit labeling of image content are shown in C.2.
5.3 Audio Content Explicit Labeling
Methods for explicit labeling of audio content are as follows.
a) Explicit labeling of audio content should be in the form of voice labels or audio rhythm labels.
b) Voice labeling should incorporate the following elements:
1) Artificial Intelligence Element: Include "Artificial Intelligence" or "AI" to indicate the use of AI technology.
2) Generated/Synthetic Element: Include "Generated" and/or "Synthetic" to indicate that the content creation method involves generation and/or synthesis.
c) Audio rhythm labels should be in the rhythm of "short long short short".
Note 1: The "short long short short" rhythm is the Morse code representation of "AI".
d) Explicit labeling of audio content should be located in one or more of the following positions:
1) At the beginning of the audio;
2) At the end of the audio
3) At an appropriate position in the middle of the audio.
Note 2: The beginning of the audio is before the start of the generated and synthesized audio content. The end of the audio is after the end of the generated and synthesized audio content.
Note 3: In high-frequency audio interaction scenarios such as smart voice assistants, smart customer service, and smart navigation, the beginning and end of the audio refer to the beginning and end of a round of interaction.
e) Voice labels should be at a normal speaking rate.
Note 4: The normal speaking rate in Chinese is approximately 120-160 characters per minute.
f) Rhythm labels should be clear and distinguishable.
Examples of explicit labeling of audio content are shown in C.3.5.4 Video Content Explicit Labeling
5.4 Video Content Explicit Labeling
Methods for explicit labeling of video content are as follows.
a) Explicit labeling of video content should be in the form of text prompts.
b) Explicit labeling of video content should simultaneously include the following elements:
1) Artificial intelligence element: including "artificial intelligence" or "AI", indicating the use of artificial intelligence technology;
2) Generation and synthesis element: including "generation" and/or "synthesis", indicating the content creation method as generation and/or synthesis.
c) Explicit labeling of video content should be located at the starting frame of the video, and may be located at the end of the video and at appropriate positions in the middle.
d) Explicit labeling of video content should be located at the edge or corner of the video frame.
e) The font of the explicit labeling of video content should be clear and distinguishable.
f) The height of the text in the explicit labeling of video content should not be less than 5% of the length of the shortest side of the frame.
g) Under normal video playback speed, the duration of the explicit labeling of video content should be no less than 2 seconds.
Examples of explicit labeling of video content are shown in C.4.
5.5 Virtual Scene Explicit Labeling
Methods for explicit labeling of virtual scenes are as follows.
a) Explicit labeling of virtual scenes should be in the form of text prompts.
b) Explicit labeling of virtual scenes should simultaneously include the following elements:
1) Artificial intelligence element: including "artificial intelligence" or "AI", indicating the use of artificial intelligence technology;
2) Generation and synthesis element: including "generation" and/or "synthesis", indicating the content creation method as generation and/or synthesis.
c) Explicit labeling of virtual scenes should be located at the starting frame of the virtual scene, and may be located at appropriate positions during the continuous service of the virtual scene.
d) Explicit labeling of virtual scenes located at the starting frame should be at the edge or corner of the frame.
e) The font of the explicit labeling of virtual scenes should be clear and distinguishable.
f) The height of the text in the explicit labeling of virtual scenes should not be less than 5% of the length of the shortest side of the frame.
5.6 Interaction Scenario Interface Explicit Labeling
Methods for explicit labeling of interaction scenario interfaces are as follows.
a) Explicit labeling of interaction scenario interfaces should be in the form of text prompts.
b) Explicit labeling of interaction scenario interfaces should simultaneously include the following elements:
1) Artificial intelligence element: including "artificial intelligence" or "AI", indicating the use of artificial intelligence technology;
2) Generation and synthesis element: including "generation" and/or "synthesis", indicating the content creation method as generation and/or synthesis.
c) Explicit labeling of interaction scenario interfaces should be implemented in one or more of the following ways:
1) Continuously displaying prompt text near the content;
2) Continuously displaying prompt text at the top, bottom, background, etc., of the interaction scenario interface.
d) The font and color of the explicit labeling of interaction scenario interfaces should be clear and distinguishable.
Examples of explicit labeling of interaction scenario interfaces are shown in Appendix D.
6. Implicit Labeling
6.1 Implicit Labeling in File Metadata
Methods for implicit labeling in file metadata are as follows.
a) Implicit labels should include the following elements:
1) Generation and synthesis tag element: information on the artificial intelligence generation and synthesis attributes of the content;
2) Generation and synthesis service provider element: name or code of the generation and synthesis service provider;
3) Content production number element: unique number assigned by the generation and synthesis service provider to the content;
4) Content dissemination service provider element: name or code of the content dissemination service provider;
5) Content dissemination number element: unique number assigned by the content dissemination service provider to the content.
b) The format of implicit labeling in file metadata should comply with Appendix E.
c) In the file of content generated by artificial intelligence, only one copy of the implicit label in file metadata should be retained. Examples of implicit labeling in file metadata are shown in Appendix F.
6.2 Implicit Labeling in Content
Implicit labeling in content is permitted in the form of digital watermarks, etc.
Appendix A (Informative) Labeling Methods
A.1 Explicit Labeling
Explicit labeling includes content explicit labeling, interaction scenario interface explicit labeling, etc.
A.2 Implicit Labeling
Implicit labeling in content is digital watermarks, etc., added in the data of content generated by artificial intelligence, and not required in this document.
Appendix B (Informative)
Typical Application Scenarios for Explicit Labeling
Typical application scenarios for explicit labeling include services that may cause the public to be confused or misled:
a) Smart dialogue, smart writing, and other services that simulate natural persons to generate or edit text;
b) Synthetic voice, voice imitation, and other voice generation or editing services that significantly alter personal identity characteristics;
c) Face generation, face swapping, and other services that generate or edit images and videos of people or significantly alter personal identity characteristics;
d) Face manipulation, posture manipulation, and other services that generate or edit images and videos of people or significantly alter personal identity characteristics;
e) Immersive realistic scene generation or editing services;
f) Text-to-image and other image content generation services;
g) Music creation and other audio content generation services;
h) Text-to-video, image-to-video, and other video content generation services;
i) Other services with functions to generate or significantly alter information content.Appendix C (Informative)
Appendix C (Informative)
Examples of Content Explicit Labeling
C.1 Text Content Explicit Labeling
C.1.1 Text Form
Example of text-form explicit labeling located at the end of the text using the prompt text "AI generation", as shown in Figure C.1.
Figure C.1 Example of text-form content explicit labeling located at the beginning of the text
Example of text-form explicit labeling located at the end of the text using the prompt text "AI generation", as shown in Figure C.2.
Figure C.2 Example of text-form content explicit labeling located at the end of the text
C.1.2 Superscript Form
Example of superscript-form explicit labeling located at the beginning of the text using the "AI" prompt superscript, as shown in Figure C.3.
Figure C.3 Example of superscript-form content explicit labeling located at the beginning of the text
Example of superscript-form explicit labeling located at the end of the text using the "AI" prompt superscript, as shown in Figure C.4.
Figure C.4 Example of superscript-form content explicit labeling located at the end of the text
C.2 Image Content Explicit Labeling
Example of image content explicit labeling located at the lower right corner of the image using the prompt text "generated and synthesized by artificial intelligence", as shown in Figure C.5.
Figure C.5 Example of image content explicit labeling located at the lower right corner of the image
C.3 Audio Content Explicit Labeling
Example of adding voice or audio rhythm labels at the beginning of the audio, as shown in Figure C.6.
Figure C.6 Example of voice or audio rhythm labels located at the beginning of the audio
C.4 Video Content Explicit Labeling
Example of video content explicit labeling located at the lower right corner of the starting frame of the video using the prompt text "generated and synthesized by artificial intelligence", as shown in Figure C.7.
Figure C.7 Example of video content explicit labeling located at the lower right corner of the starting frame of the video
Appendix D (Informative)
Examples of Explicit Labeling in Interaction Scenario Interfaces
D.1 Continuous Display Near Content
D.1.1 Near Content Display Area
Example of explicit labeling in interaction scenario interfaces located near the text content display area using the prompt text "AI generation", as shown in Figure D.1.
Figure D.1 Example of explicit labeling in interaction scenario interfaces located near the text content display area
D.1.2 Near Audio Playback Area
Example of explicit labeling in interaction scenario interfaces located near the audio playback area using the prompt text "AI generation", as shown in Figure D.2.
Figure D.2 Example of explicit labeling in interaction scenario
interfaces located near the audio playback area
D.2 Continuous Display in Appropriate Positions of Interaction Scenario Interfaces
D.2.1 At the Bottom of Interaction Scenario Interfaces
Example of explicit labeling in interaction scenario interfaceslocated at the bottom using the prompt text "generated and synthesized by artificial intelligence", as shown in Figure D.3.
Figure D.3 Example of explicit labeling in interaction scenario interfaces located at the bottom
D.2.2 In the Background of Interaction Scenario Interfaces
Example of explicit labeling in interaction scenario interfaces located in the background using the prompt text "generated and synthesized by artificial intelligence", as shown in Figure D.4.
Figure D.4 Example of explicit labeling in interaction scenario interfaces located in the background.
Appendix E (Normative)
Format of Implicit Labeling in File Metadata
Requirements for the format of implicit labeling in file metadata are as follows.
a) Add an implicit labeling extension field in the file metadata, with the field name or keyword including "AIGC".
b) The value of the implicit labeling extension field should be a string in the following format.
{"AIGC":
{"Label":"value1","ContentProducer":"value2","ProduceID":" value
","ReservedCode1":"value4","ContentPropagator":"value5","Propag ateID": " value6 ","ReservedCode2":"value7 "}}
Note 1: When the content generation and synthesis service provider first writes the implicit label in file metadata for content generated and synthesized by artificial intelligence, the content dissemination service provider element is consistent with the content generation and synthesis service provider element, and the content dissemination number element is consistent with the content production number element.
c) The generation and synthesis label element is represented by Label, with the value being value1, and should meet the following requirements.
1) Stores information on whether the content is, may be, or is suspected to be generated and synthesized by artificial intelligence: if it is generated and synthesized by artificial intelligence, the value of value1 is 1; if it may be generated and synthesized by artificial intelligence, the value is 2; if it is suspected to be generated and synthesized by artificial intelligence, the value is 3.
2) The type is string.
d) The generation and synthesis service provider element is represented by ContentProducer, with the value being value2, and should meet the following requirements:
1) Stores the name or code of the generation and synthesis service provider;
2) The type is string.
e) The content production number element is represented by ProduceID, with the value being value3, and should meet the following requirements:
1) Stores the unique number assigned by the generation and synthesis service provider to the content;
2) The type is string.
f) The reserved field 1 is represented by ReservedCode1, with the value being value4, and the requirements are as follows:
1) Can store information used by the generation and synthesis service provider to independently carry out security protection and ensure the integrity of the content and labels;
2) The type should be string.
Note 2: An example of a generation and synthesis service provider using reserved field 1 for security protection of implicit labels in file metadata is shown in Appendix F, F.4.
g) The content dissemination service provider element is represented by ContentPropagator, with the value being value5, and should meet the following requirements:
1) Stores the name or code of the content dissemination service provider;
2) The type is string.
h) The content dissemination number element is represented by PropagateID, with the value being value6, and should meet the following requirements:
1) Stores the unique number assigned by the content dissemination service provider to the content;
2) The type is string.
i) The reserved field 2 is represented by ReservedCode2, with the value being value7, and the requirements are as follows:
1) Can store information used by the content dissemination service provider to independently carry out security protection and ensure the integrity of the content and labels;
2) The type should be string.
j) The values of the elements in c) to i) should mainly consist of characters with code positions 0x21, 0x23 to 0x5B, 0x5D to 0x7E in GB18030—2022, as well as ".
Note 3: Code positions 0x21, 0x23 to 0x5B, 0x5D to 0x7E include all single-byte encoded characters except " and \, and " is an escape character used to represent ".
Appendix F (Informative)
Examples of Implicit Labeling in File Metadata
F.1 Example of Implicit Labeling in Image File Metadata
Example of implicit labeling in the metadata of an image file generated and synthesized by artificial intelligence, as shown in the "AIGC" part of Figure F.1.
Figure F.1 Example of implicit labeling in image file metadata
F.2 Example of Implicit Labeling in Audio File Metadata
Example of implicit labeling in the metadata of an audio file that may be generated and synthesized by artificial intelligence, as shown in the "AIGC" part of Figure F.2.
Figure F.2 Example of implicit labeling in audio file metadata
F.1 Example of Implicit Labeling in Video File Metadata
Example of implicit labeling in the metadata of a video file suspected to be generated and synthesized by artificial intelligence, as shown in the "AIGC" part of Figure F.3.
Figure F.3 Example of implicit labeling in video file metadata
F.1 Example of Security Protection Using Digital Signatures
An example of a generation and synthesis service provider using a hash algorithm to digitally sign file metadata information and storing the result in reserved field 1 is as follows. "ReservedCode1":"e862483430d978cbf828b8b24296ef9328d843a0"