China: TC260 requests comments on practice guidelines for content identification method for generative AI
On August 8, 2023, the National Information Security Standardization Technical Committee (TC260) requested public comments on the Practice Guidelines for Network Security Standards - Content Identification Method for Generative Artificial Intelligence Services.
Specifically, the Guidelines outline methods for content identification by generative artificial intelligence (AI) services, and apply to the use of generative AI technology to provide the public with generated text, graphics, audio, video, and other content.
The Guidelines highlight that generative AI services may add an explicit watermark, which will not affect the normal use of content, but can still be clearly distinguished, with the transparency of the watermark set to 90%. Likewise, generative AI services may modify pictures, audio, and video content through an implicit watermark that cannot be perceived by individuals, but can be extracted through technical means.
Regarding the identification of AI-generated content, the watermark should be provided in a display area or through continuously displaying the prompt text through the input information area. For example, the Guidelines note that an explicit watermark with the text 'generated by artificial intelligence' or 'generated by AI' should be used. When pictures and videos generated by AI are used, the prompt text must be added as a form of identification, and placed in the four corners of the screen, and the text height not less than 20 pixels. The content of the prompt text must contain at least 'AI Generate,' among other examples.
On the other hand, when AI is used to generate pictures, audio, and video, an implicit watermark should be added. The implicit watermark must contain at least the name of the service provider.
Finally, when pictures, audio, and video generated by AI are output in the form of files, the watermark must be included in the file, and an extension field added to the metadata of the file for identification.
Public comments can be submitted via email to [email protected] until August 11, 2023.
Update: September 5, 2023
TC260 publishes finalized practice guidelines for content identification methods for generative AI
On August 25, 2023, the TC260 published its finalized Practice Guidelines for Network Security Standards - Content Identification Method for Generative Artificial Intelligence Services. In line with the draft practice guidelines, the finalized version provides guidance on content identification for generative AI services and applies to generative AI service providers that use generative AI technology to provide the public with generated text, graphics, audio, video, and other content.