On 1 September 2025 the Measures for Labelling AI-Generated Synthetic Content took effect in China. The measures were jointly issued by the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration, with a supporting mandatory national standard taking effect at the same time. For AI service providers and content platforms, this is a clear milestone: content labelling has moved from an industry expectation to a legally binding duty.

The central risk of synthetic content is not whether it looks convincing, but whether people can tell it apart. When a video, image or voice recording can be fabricated at low cost, the public’s basic ability to judge the authenticity of information erodes, and the consequences reach reputation, portrait rights, copyright and market order.

Before the measures, labelling requirements came mainly from the Provisions on the Administration of Deep Synthesis of Internet Information Services, which were relatively principle-based. The measures break “shall label” into two operational categories — explicit labelling and implicit labelling — and allocate responsibility by stage.

Two kinds of labelling, three sets of duties

Explicit labelling targets user perception. It requires a text or graphic notice placed where users and sharers can readily perceive the AI nature of the content.

Implicit labelling targets technical traceability. It requires generation attributes and provider identifiers to be written into file metadata, leaving a technical interface for later verification and platform handling.

Responsibility is distributed across three stages: service providers must label at the generation and output stage; content platforms must verify and warn, and act on content that appears unlabelled; users may not maliciously delete, alter or forge labels. This turns labelling from a single act into a chain of duties. See our note on labelling enforcement in practice.

How it connects to existing regimes

The measures are not an isolated document. They work alongside the deep synthesis algorithm filing regime: filing answers whether a service was registered before launch, labelling answers whether output can be identified afterwards. See our note on batch-based deep synthesis filing.

They also connect to the generative AI security standard, which converts corpus, content and labelling requirements into testable technical clauses — full labelling coverage is one of them. See our note on the generative AI security standard.

Labelling duties also interact with user verification obligations: when a platform handles user uploads, it must assess whether content is synthetic, which creates new verification workflows. See our note on user verification duties for AI-generated content.

Four practical footholds

First, make labelling a default. Text, image, audio and video output paths should all carry explicit and implicit labelling, avoiding partial gaps.

Second, retain labelling records. Labelling acts should be logged so a company can show it met its duty in a regulatory inquiry or platform dispute.

Third, tighten the user terms. Prohibit users from removing or altering labels, and back the prohibition with a handling mechanism. See our algorithm filing service.

Fourth, build a reporting and handling loop. Establish discovery, verification and handling steps for unlabelled content, rather than relying on declarations alone. See our note on AI misuse reporting channels.

The measures mark AI content governance moving from setting the rules to keeping them. For companies, building labelling into product defaults is cheaper than explaining afterwards.

This article is based on the Measures for Labelling AI-Generated Synthetic Content and the supporting national standard, and on public information. It is for general reference only and does not constitute legal advice.