Labelling duties for deep synthesis services have been in force for some time, and 2024 enforcement practice has turned them from a line of notice into an auditable compliance item. Dedicated labelling rules are still being drafted, and public discussion has settled on a distinction between explicit and implicit labels. The rules are not yet in force, but the direction is clear: labelling is not only notice to users, it is infrastructure for tracing content.
What each type of label solves
Regulatory discussion and industry practice usually divide labels into two kinds. An explicit label lets the user perceive that content was AI-generated, through text notice, voice notice or an on-screen marker. An implicit label is written into file metadata or a digital watermark for platforms and regulators to read in the background. The first addresses awareness; the second addresses traceability.
Neither substitutes for the other. With only explicit labels, content becomes hard to trace once screenshotted or re-uploaded. With only implicit labels, users watching the content do not know its origin. Platforms with long distribution chains need both.
Dividing duties between three parties
Tool and service providers add labels at the point of generation; that is the source duty. Distribution platforms verify, display and retain label information at the point of distribution. Users should declare synthetic content when uploading or publishing. Duties stack rather than transfer: a tool’s label does not relieve the platform of verification, and a platform’s notice does not relieve the user of disclosure.
Disagreement usually arises over whose technical solution should carry the load. The tool says metadata was written; the platform says it was lost on upload. The platform asks users to label; users say the tool should do it automatically. Writing the technical interface and the responsibility boundary into the cooperation agreement is more useful than arguing after a dispute.
The practical difficulty of platform verification
Metadata is frequently lost in transcoding, screenshots and re-editing, so verification based on metadata alone will miss cases. Verification based on content detection alone is hard to keep within an acceptable error rate. A layered approach works better: read metadata from direct uploads, run detection on suspected content, and add human review for high-risk scenarios.
Verification also needs appeal and restoration. A false positive affects creators directly, so the platform should explain the basis, provide an appeal channel and record the outcome. Without that, the labelling regime transfers cost to content producers.
A minimum action list
Map the generation points: which products, features and interfaces produce synthetic content. Add explicit and implicit labels at output. Retain label versions and writing logs. State the disclosure duty in user terms. Agree with upstream and downstream parties on how labels are passed on and what happens if they are lost. Establish a complaint and handling process.
For content distributed across borders, check whether labelling requirements are compatible with local rules. Label fields, language and presentation differ between jurisdictions, so a unified technical solution needs room for configuration.
For labelling design and launch planning, see our AIGC labelling service; for filing and materials, see algorithm filing; for infringement and complaints, see AI infringement.
This article is based on China’s Provisions on the Administration of Deep Synthesis of Internet Information Services and on public information available in 2024. It is for reference only and does not constitute legal advice.