Labeling of AI-generated and synthesized content
Text content needs a written prompt or a generic symbol prompt at a suitable position at the start, the end or the middle, and may also carry a conspicuous notice in the interactive interface or around the text; audio needs a voice prompt or an audio rhythm prompt at a suitable position at the start, the end or the middle, and may also carry a conspicuous notice in the interactive interface. Either approach may be chosen, or both together. In addition, where download, copy or export functions are offered, the file must contain a compliant explicit label.
Article 4 limits the premise for explicit labelling: the generative service offered by the service provider falls within the circumstances in Article 17, paragraph 1 of the Provisions on the Administration of Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定). Where that premise is met, explicit labels shall be added as follows.
For text, (1) add a text prompt or a generic symbol prompt and similar labels at a suitable position at the start, the end or the middle of the text, or add a conspicuous notice in the interactive interface or around the text.
For audio, (2) add a voice prompt or an audio rhythm prompt and similar labels at a suitable position at the start, the end or the middle of the audio, or add a conspicuous notice in the interactive interface.
The final paragraph of the same article: where a service provider offers functions such as downloading, copying or exporting generated synthetic content, it shall ensure that the file contains a compliant explicit label.
Text products: add one clear notice at the start or the end of the body (for example stating that the content is generated by artificial intelligence), or place a symbol notice at a fixed position around the text, inside the first screen the user sees.
Audio products: insert a voice prompt at the start or the end, or use a rhythm cue; at the same time give a conspicuous notice in the playback interface, so that a user who only listens and does not look is not left with no perception at all.
Interactive scenarios (chat, voice customer service, generation tool interfaces): display the notice in a fixed area of the interface permanently, rather than a one-off pop-up on first use that then disappears.
Export paths: write the requirement that the explicit label survives export as a test case, and re-test after transcoding, compression and concatenation.
Putting the notice only at the bottom of the page or in the help documentation, where ordinary users never see it.
Using light grey small text with almost no contrast against the background, which is easily found not conspicuous.
Adding the notice only at the end of the audio: a user who drops out halfway gets no label at all.
Making the notice a pop-up that can be dismissed with one click, leaving the page with no label afterwards.
Labelling only the web version, while the app, the mini program and exported files have nothing.
Q: Do I have to add notices at the start, the end and the middle?
A: The clause uses or, so choosing one of them at a suitable position is enough — but the test remains whether the user can clearly perceive it.
Q: For AI dubbing used in internal corporate training, is a voice prompt required?
A: First assess whether that service falls within the premise of Article 4, then decide the labelling method; treat services offered to the public strictly, and follow your local cyberspace administration on borderline cases.
This entry belongs to the “China AI Compliance and Filing” cluster (42 entries in total): View all · Answers home
Who we are: a knowledge base and AI customer service team serving small and micro businesses, delivering implementations that take AI customer service through the national standard and the filing process.
Content on this site is compiled from publicly released regulatory texts for enterprises' own self-check reference; the specific filing requirements of the local cyberspace administration authorities prevail, and this is not legal advice.
This page is generated from the China AI compliance and filing MCP corpus — agents can call the same data directly.