Improving the Labeling Methods for AI-Generated Synthetic Content to Build a Healthy Information Network Ecosystem

2025-02-06

Reposted from the Cyberspace Administration of China

 

With the rapid development of generative artificial intelligence (AI) technologies, the use of AI to generate or synthesize text, images, music, videos, 3D models, and other digital content has become a primary method of content production in cyberspace. This trend continues to inject strong momentum into the digital economy, industrial development, and even societal transformation. However, as the barriers to content generation and synthesis continue to lower, the misuse of these new technologies poses greater challenges for cyberspace governance. Currently, countries such as China, the United States, and those in the European Union are exploring the incorporation of content disclosure and labeling obligations into legislation. Meanwhile, numerous internet companies worldwide are gradually implementing labeling embedding and management measures. Content labeling has become a global practice in combating false and harmful information. In response to this trend and industry demand, the Cyberspace Administration of China, along with other relevant departments, has timely released the Measures on the labeling method for AI-Generated Synthetic Content (hereinafter referred to as the Measures) and the accompanying national standard Cybersecurity Technology - Labeling method for content generated by artificial intelligence (hereinafter referred to as the Labeling method). These documents comprehensively and multi-dimensionally regulate the requirements and methods for identifying AI-generated content, marking an important step in improving China's AI governance framework.

1. Problem-Oriented Approach: Addressing New Demands and Risks in AI-Generated Synthetic Content

Rooted in the healthy development and regulated application of generative AI, the Measures adopt a problem-oriented approach, offering systematic and innovative institutional designs to fully address the requirements for identifying AI-generated synthetic content.

Firstly, proposing a Chinese solution for identifying synthetic content.

In response to the disruptive risks and challenges posed by AI-generated content, the international community is also exploring methods and pathways for content labeling. The EU AI Act, which came into effect in August 2024, mandates transparency obligations for AI-generated synthetic content, requiring platforms to add mandatory watermarks. In August 2024, California, USA, released the California Digital Content Provenance Standards. which requires the embedding of traceable data in AI-generated content to ensure transparency of content sources. In July 2024, the International Telecommunication Union (ITU) published the Detecting deepfakes and generative AI: Report on standards for AI watermarking and multimedia authenticity workshop, which preliminarily discusses collaborative arrangements for detecting deepfake and generative AI content, including content labeling. Grounded in China's developmental practices, the Measures take a long-term, forward-looking, and rapid-action approach, pioneering the enhancement of the comprehensive governance mechanism for generative AI through standards. This effort strongly supports the governance philosophy of "technology for good" and contributes a Chinese solution to global AI governance that balances development and security.

Secondly, strengthening the standardization of content labeling by service providers.

With the increasing accessibility and popularity of AI-generated synthetic content tools, the threshold for applying these technologies has significantly lowered. As early as 2022, the Cyberspace Administration of China and other departments issued the Regulations on the Administration of Deep Synthesis in Internet Information Services (hereinafter referred to as the Regulations), which clearly stipulate that deep synthesis service providers must prominently label generated or edited content in reasonable locations or areas to inform the public of its synthetic nature. However, the Regulations primarily specify the situations requiring labeling but do not provide detailed requirements for the placement, font, or expression of such labels. As a result, various enterprises have encountered issues such as insufficiently noticeable text labels, missing or weakly indicated audio and video labels, and sparse or difficult-to-verify implicit labels, posing challenges for regulatory enforcement. The introduction of the Measures is a necessary step to further advance the labeling of synthetic content, standardize the management of deep synthesis in internet information services, address prominent issues in content labeling, and promote the governance of the information network ecosystem, with significant and far-reaching implications.

Thirdly, further improving the accountability system for generative AI regulation.  

The generation and synthesis of content by generative AI involve multiple types of entities. Clearly defining the obligations of each entity helps to solidify responsibility, facilitate technical detection, and implement precise accountability. The Measures not only define the basic concepts of AI-generated synthetic content but also specify specific labeling requirements for service providers, dissemination platforms, application service distribution platforms, and users, providing clear directions for compliance. In doing so, the Measures supplement the existing accountability system, elaborating on the behavioral norms for deep synthesis service providers and technical supporters proposed in the Regulations, and offering a valuable example for promoting the regulated development of new technologies and applications.

2. Clear Requirements: Promoting Fine-Grained Governance of AI-Generated Synthetic Content

AI-generated synthetic content encompasses not only text, images, audio, and video but also provides more implementation solutions in interactive and immersive scenarios such as the "metaverse." The Labeling method comprehensively cover AI-generated synthetic content and further clarify the requirements and operational guidelines of the Measures.

Firstly, distinguishing between explicit and implicit labeling requirements.

To meet the needs of both perceptible and imperceptible labeling method, the Labeling method clearly outline the key elements for each. For explicit labeling, requirements are proposed around text elements, placement, font, color, audio rhythm, video pixels, and interactive prompts. For metadata and implicit content labeling, key guidance is provided on elements such as synthetic labels, service provider information, and content numbering.

Secondly, listing typical application scenarios for explicit labeling.

The Labeling method further refine the application scenarios for explicit labeling through enumeration. In addition to the five types of prominent labeling scenarios mentioned in the Regulations, three new typical application scenarios are added, including image generation services (e.g., text-to-image), audio generation services (e.g., music), and video generation services (e.g., text-to-video, image-to-video). These three new scenarios represent rapidly emerging forms of synthetic content since 2022, particularly with the powerful potential demonstrated by ChatGPT and Sora in video content generation. Including these scenarios in the Labeling method facilitates targeted and forward-looking regulation, aligning with the technical principles and industrial needs of generative AI.

Thirdly, standardizing metadata implicit labeling format requirements.

Previously, implicit Labeling method for synthetic content varied among enterprises of different scales and technical levels, and the label information in these labeling was inconsistent. This inconsistency posed challenges for regulatory bodies and content platforms in recognizing implicit labels. To address this issue, the Labeling method provide detailed explanations for extending field keywords in implicit labeling, including the labels, numbering, and reserved field values that should be included. This offers an effective measure for the unified, stable, and sustained implementation of implicit labeling requirements across the industry.

Finally, demonstrating examples of explicit and implicit labeling requirements.

The Labeling method use image examples to illustrate various explicit and implicit labeling requirements, helping enterprises intuitively understand the standard requirements and avoid ambiguity or misinterpretation caused by textual descriptions. This approach promotes industry-wide collaborative implementation and enhances governance efficiency.

3. Upholding Integrity While Innovating: Balancing Content Generation Service Management with Technological Innovation

AI-generated synthetic technologies should be used in a regulated and accelerated manner under the premise of safety, with high-level governance ensuring high-quality development. The Measures and Labeling method fully consider the needs of both "use" and "management," minimizing the impact of content labeling on user experience while encouraging enterprises to voluntarily display labeling labels.

Firstly, providing multiple Labeling method to adapt to diverse product forms.

The Labeling method innovatively propose corner mark forms for explicit text labeling, requiring only the addition of an "AI" corner mark at the end of the content to efficiently meet requirements. For audio content, the Labeling method introduce audio rhythm labeling for the first time, suggesting the use of Morse code with a "short-long-short-short" rhythm at appropriate positions in the audio to achieve the desired prompt effect. These Labeling methods not only facilitate the creation of diverse product forms for enterprises but also ensure user experience and cognitive effectiveness, achieving regulatory goals while providing effective safeguards for technological development.

Secondly, dividing responsibilities among different entities and refining the application of the safe harbor principle.

Although the Measures primarily impose the responsibility of adding labels on service providers and users who upload content, they do not apply a "one-size-fits-all" safe harbor rule to dissemination platforms and application distribution platforms. Instead, the Measures adopt a more refined approach, stipulating that dissemination platforms should verify the extended fields in file metadata, prompt users about synthetic content information, and add dissemination element information to file metadata. Application service distribution platforms should verify whether service providers have provided the required synthetic content labeling functions before app listing or launch. This approach helps appropriately adjust the distribution of obligations among service providers, users, and dissemination platforms, preventing related platforms from "hiding" within the deep synthesis regulatory system. It truly achieves a collaborative governance mechanism, effectively maintaining a healthy and positive information network ecosystem.