Interpretation of the Draft Method for Identifying Synthetic Content Generated by Artificial Intelligence

LABEL： AI , Compliance business , Digital economy ,

On September 14, 2024, the State Internet Information Office released the Methods for the Identification of Artificial Intelligence Generated Synthetic Content (Draft for Comments) (hereinafter referred to as "AIGC Content Identification Methods"), which aims to standardize the artificial intelligence generated synthetic content identification. At the same time, as a supporting national standard for the "AIGC Content Identification Method", the "Network Security Technology Artificial Intelligence Generated Composite Content Identification Method (Draft for Comments)" (hereinafter referred to as the "Identification Method") was also announced on the same day.

Prior to the release of the AIGC Content Identification Method, China's Administrative Provisions on Algorithm Recommendation of Internet Information Services (hereinafter referred to as the "Algorithm Recommendation Rules"), Administrative Provisions on Deep Synthesis of Internet Information Services (hereinafter referred to as the "Deep Synthesis Rules") and Interim Measures for the Management of Generative Artificial Intelligence Services (hereinafter referred to as the "AIGC Method") have proposed the general provisions on the algorithm generation of synthetic information and the performance of identification obligations on deep synthesis of information content (including pictures, videos, etc.). We understand that the "AIGC Content Identification Measures", based on the aforementioned regulations, clarify the governance object of artificial intelligence generated and synthesized content, and propose further detailed provisions on the identification obligation subject, scope of application, identification type, identification deployment scenario and deployment method, etc., in order to promote the implementation of the obligation of artificial intelligence generated and synthesized content identification in practice, promote the healthy development of artificial intelligence, protect the legitimate rights and interests of citizens, legal persons and other organizations, and safeguard social public interests. This article will briefly introduce the main content of the "AIGC Content Identification Method".
1、 Main content of the AIGC Content Identification Method
（1） Clearly define the applicable objects and scope

According to Article 2 of the AIGC Content Identification Measures, the core entity responsible for fulfilling the identification obligation is the network information service provider (hereinafter referred to as the "service provider") that generates synthetic content using artificial intelligence. This regulation complies with the provisions of the "Deep Synthesis Regulations" and the "AIGC Measures" regarding the obligation of identifying the subject of generating and synthesizing content, that is, the subject of fulfilling the obligation of content identification should be organizations or individuals who provide generating and synthesizing content services using deep synthesis technology and generative artificial intelligence technology, and the direct subject of obligation does not currently include organizations or individuals who only provide technical support for artificial intelligence generated and synthesized content.

Secondly, Article 2 of the "AIGC Content Identification Measures" clearly stipulates that those who develop and apply artificial intelligence generation and synthesis technology but do not provide services to the domestic public are not subject to the provisions of this measure. The legislative logic of this scope of application is consistent with the AIGC Measures, and is an extraterritorial application clause established on the basis of the provisions on territorial jurisdiction effectiveness in the Algorithm Recommendation Regulations and the Deep Synthesis Regulations. We understand that in practice, whether related enterprises provide artificial intelligence generated composite content services in China through embedded integration or encapsulation in products and services through application programming interfaces, there is a possibility of applying the AIGC Content Identification Method. Therefore, relevant enterprises should fully consider the legality and compliance of such overseas services under Chinese law..
（2） Distinguish the types, usage scenarios, and deployment methods of identification

Previously, the "Algorithm Recommendation Regulations" and "Deep Synthesis Regulations" only imposed the obligation to prominently label the content of algorithm generated synthesized information and related deep synthesized information in reasonable locations and regions, but did not further explain the specific types of labeling, the specific orientation of "reasonable locations and regions", and the degree of recognition of "saliency".

The "AIGC Content Identification Measures" issued this time respond to the aforementioned content that needs to be clarified: firstly, based on whether users can perceive it clearly as the judgment criterion, the measures clearly propose two types of artificial intelligence generated composite content identification, namely "explicit identification" and "implicit identification". Secondly, the "AIGC Content Identification Measures" clarify the scenarios in which the two types of identification should be applied, establishing a direct correspondence between the "prominent identification" and "explicit identification" proposed in Article 17 of the "Deep Synthesis Regulations". Finally, based on the carriers and forms of synthetic content generated by different artificial intelligence (such as text, audio, images, etc.), the "AIGC Content Identification Method" lists and explains the deployment locations of various explicit identifiers and the information elements that implicit identifiers should include. We will provide a visual summary of this in the form of the following table:

           

At the same time, the "Identification Method" published on the same day as the "AIGC Content Identification Method" provides clearer guidance on the methods of explicit and implicit identification. The "Identification Method" is applicable to standardize the identification activities carried out by providers of synthetic services and content dissemination services for artificial intelligence generated synthetic content. It provides more specific regulations on the identification methods for various types of artificial intelligence generated content and interactive scene interfaces from the aspects of application scenarios, identification methods, and identification examples. In terms of explicit identification, the "Identification Method" specifies the form, location, and elements of identification in different application scenarios (including text content, image content, audio content, video content, etc.). In terms of implicit identification, the "Identification Method" specifies that the implementation methods of implicit identification include metadata implicit identification, content implicit identification, etc., and provides specific explanations on the elements and formats of metadata implicit identification (see the summary in the table below).

           
（3） Verification and identification obligations for service providers of newly added network information content dissemination platforms

As mentioned earlier, the verification and identification obligations stipulated in the "Algorithm Recommendation Regulations," "Deep Synthesis Regulations," and "AIGC Measures" were mainly imposed on network information service providers who generate synthetic content using artificial intelligence. However, Article 6 of the AIGC Content Identification Measures includes service providers of network information content dissemination platform services (hereinafter referred to as "platform service providers") as the obligation subjects for verification and identification. This method standardizes the obligations of platform service providers from multiple dimensions, including:

Verification obligation: Platform service providers shall verify whether the metadata of the document contains implicit identifiers to identify the generated composite content (Article 6, Paragraph 1).

Obligation to Identify: Based on verification results, user statements, and proactive detection, platform service providers should add prominent warning signs around the published content to remind users of the nature of the content (whether it is indeed, may be, or suspected of generating synthetic content) (Article 6, paragraphs 2 and 3).

Obligation to improve metadata: For relevant content, attribute information for generating composite content and dissemination element information should be added to the file metadata to enhance the traceability of the content (Article 6, Paragraph 4).

Function provision and user education obligation: Provide necessary identification functions to remind users to actively declare whether the published content includes generated composite content (Article 6, Paragraph 5).

Firstly, considering the effectiveness of the identification, this method is conducive to forming a more comprehensive regulatory system. Platform service providers directly provide online information services to the public, which directly determines the information content that users come into contact with and plays an important role in the information dissemination chain. Meanwhile, platform service providers typically possess more advanced technological means and resources to detect and identify generated composite content, such as metadata verification, content analysis, etc., in order to fulfill their verification and identification obligations. Including platform service providers in the scope of responsibility will help form a full chain, multi-level regulatory system, with clear responsible parties in every link from content generation, processing to dissemination.

Secondly, from the perspective of legislative ideas, increasing the verification and labeling obligations of platform service providers is also in line with the previous legislative ideas. The Provisions on Security Assessment of Internet Information Services with Public Opinion Attribute or Social Mobilization Capability clearly stipulates that "if the use of new technologies and new applications causes significant changes in the functional attributes, technology realization methods, basic resource allocation, etc. of information services, leading to significant changes in the public opinion attribute or social mobilization capability," security assessment should be conducted. The Administrative Provisions on Algorithm Recommendation of Internet Information Services also requires that algorithm recommendation service providers with the attribute of public opinion or social mobilization ability perform the obligation of algorithm filing. Platform service providers are mainly social media, content sharing platforms, forums, blogs, etc., and most of them can be identified as Internet information service providers with public opinion attributes or social mobilization capabilities. Due to its important position and influence in information dissemination, laws and regulations have set stricter regulatory measures and requirements to prevent potential risks, including conducting security assessments, algorithm filing, etc. The new content verification and identification obligations reflect the unified legal supervision of key Internet service providers, that is, to maintain the security and order of cyberspace by strengthening the supervision of the core links of information dissemination.

Finally, considering the principle of responsibility, there is a significant difference between the establishment of this regulation and the "safe harbor principle". The principle of safe haven is an important principle in Internet law, which aims to balance the responsibilities of network service providers with the needs of Internet development. Its core contents include passive neutral obligations, notification deletion mechanism and liability exemption, that is, as long as service providers do not participate in the production, selection or modification of content, and take necessary measures in a timely manner after receiving the notice, corresponding legal liabilities can be exempted. The responsibility of platform service providers emphasizes proactive responsibility, and service providers need to proactively verify, detect, and identify the generated composite content. The safe harbor principle emphasizes passive neutrality, where service providers do not need to actively monitor user content and only take action after being notified. That is to say, the responsibility of service providers requires them to shift from passive neutrality to proactive action. But we understand that the safe harbor principle still applies in the general field of content provision, except in special areas involving artificial intelligence generated and synthesized content, platform service providers need to bear higher obligations and responsibilities of care.
（4） The prerequisites and identification obligations for new users to obtain content without explicit identification

Previously, the "Regulations on Deep Synthesis" only proposed that deep synthesis service providers should provide prominent identification functions, and reminded deep synthesis service users (including users) of the obligation to perform prominent identification, as well as the obligation of no organization or individual to use technical means to delete, tamper with, or conceal deep synthesis identification. On this basis, previous regulations have not made clear requirements on whether users have an obligation to identify the generated and synthesized content they disseminate.

At present, considering that users may have practical needs for obtaining generated composite content without explicit identification and uploading generated composite content to network information content dissemination platforms, Articles 9 and 10 of the AIGC Content Identification Measures respectively propose the prerequisites for service providers to provide users with generated composite content without explicit identification, and put forward the requirements for users to upload generated composite content to service providers that provide network information content dissemination platform services. Specifically, if the user requires the service provider to provide generated composite content without explicit identification, the service provider shall fulfill the obligation to clarify the user's identification obligations and usage responsibilities through the user agreement in advance, and retain relevant logs for no less than six months. In addition, when users upload generated composite content to online information content dissemination platforms, they should actively declare and use the identification function provided by the platform to identify the generated composite content. In other words, users have the right to obtain generated composite content without explicit identification, but if users need to propagate AI generated content through online information content dissemination platforms, they still need to fulfill the obligation of identification.

As mentioned earlier, the AIGC Content Identification Measures aim to build a comprehensive and multi-level regulatory system, incorporating more responsible parties into the obligation system of artificial intelligence generated and synthesized content identification. If users only use artificial intelligence to generate content for themselves, from the perspective of convenience for users, service providers should provide them with generated composite content without explicit identification. But if users upload and spread AI generated content through online information content dissemination platforms, their role positioning shifts from service users to content disseminators, and they should annotate the AI generated content they spread and assume corresponding annotation obligations. The user identification obligation, combined with the verification and identification obligations of platform service providers as stipulated in Article 6 of the AIGC Content Identification Measures, helps to ensure that the artificial intelligence generated content that the public comes into contact with has been adequately labeled.

In addition, it is not difficult to find through an overall review of the labeling obligations that the "AIGC Content Labeling Measures" mainly regulate the content generated by artificial intelligence from a communication perspective, assigning different labeling obligations to different entities. Service providers directly generate artificial intelligence generated content and disseminate it directly to the public. For service providers, it is the starting point of the content dissemination chain and also the starting point of the labeling obligation. Therefore, the strictest labeling obligation is imposed on them. As an important platform for disseminating artificial intelligence generated content to the public, online information content dissemination platforms are prone to possessing public opinion attributes or social mobilization capabilities. Therefore, corresponding annotation obligations have been imposed on both the substantive content disseminators - users, and the formal content disseminators - platforms, jointly building a full chain, multi-level annotation obligation and responsibility system. Meanwhile, if users have not disseminated artificial intelligence content, it should not be included in the labeling obligation and responsibility system. Therefore, Article 9 of the AIGC Content Labeling Measures grants users the right to obtain content without explicit labeling.
epilogue

The promulgation of the "AIGC Content Identification Measures" marks an important step forward for China in regulating and managing the field of artificial intelligence. We understand that firstly, fulfilling the obligation of generating and synthesizing content identification has always been a difficult problem faced by service providers. The release of the "AIGC Content Identification Measures" provides a more executable implementation guide for the fulfillment of identification obligations by various responsible parties. Secondly, this method clarifies the responsibilities and obligations of service providers, online information content dissemination platforms, and users in the production and use of synthesized content, and constructs a clearer legal responsibility system. Thirdly, with the implementation of the initiative of adding explicit or implicit identification to the generated and synthesized content, the transparency of Internet information content will be significantly improved, and users can intuitively identify whether the content is generated by artificial intelligence on this basis, which helps users avoid the negative impact of false content. Finally, the application of implicit identifiers such as digital watermarks helps protect the rights of content creators and prevent unauthorized copying and distribution of content. It is worth noting that China's implementation of the "AIGC Content Labeling Measures" is also in line with the international exploration of AI generated content labeling, which helps to promote cooperation and exchange on AI governance worldwide and jointly address the challenges brought by generative AI technology.

However, at the same time, we understand that there are still some issues that need to be clarified in the 'AIGC Content Identification Method'. For example, firstly, for content suspected of being generated and synthesized, Article 6 of the "AIGC Content Identification Measures" stipulates the obligation of prompt identification and metadata improvement to be fulfilled by network information content dissemination platform service providers, that is, prompt identification should be added around the content, and attribute information and dissemination element information of generated and synthesized content should be added in the file metadata. However, the rules for the application of prompt identification and metadata improvement obligations, as well as how these obligations have priority applicability and whether they need to be applied simultaneously, need to be further clarified. Secondly, for content suspected of being generated and synthesized, considering that network information content dissemination platform service providers may only act as custodians of generated and synthesized content and have not directly or participated in the production of related artificial intelligence generated and synthesized content, there may still be room for discussion on whether the obligation to improve metadata has practical and technical feasibility for network information content dissemination platform service providers. Again, regarding the specific connotation of users' "use responsibility" through user agreements, previous legislation such as the "Deep Synthesis Regulations" and the "AIGC Measures" have not provided clear provisions. Therefore, the specific extension and boundaries of users' "use responsibility" also need to be further clarified. Finally, the "AIGC Content Identification Measures" have added a user identification obligation, but it is also worth further consideration in the "AIGC Content Identification Measures" regarding how to carry out follow-up supervision on users' fulfillment of the re identification obligation, the relevant supervision methods, and how to determine the supervisory subject.

In summary, we still look forward to the gradual improvement of the AIGC Content Identification Method, which will bring a healthier, more transparent, and responsible artificial intelligence application environment to the public.