China's AI Rules: Measures to Labeling Law

Verified directly against the Cyberspace Administration of China’s (CAC) own published notices, plus a China Briefing analysis for corroboration and CASRAI’s own NIKOLAI element pages; last checked September 2026. China does not regulate generative AI through one comprehensive statute. It regulates through two central-government administrative measures issued eighteen months apart — the 2023 Interim Measures for the Management of Generative AI Services, which governs how a generative-AI service has to operate, and the 2025 Measures for Labeling of AI-Generated Synthetic Content, which mandates that AI-generated text, images, audio, video, and virtual scenes carry both a visible label and embedded metadata — and, as of this writing, no dedicated national AI Law sits above either one. CASRAI’s broader AI regulations around the world guide gives China one paragraph; this is the deep-dive that paragraph pointed to.

The 2023 Interim Measures for Generative AI Services

The Interim Measures for the Management of Generative AI Services (生成式人工智能服务管理暂行办法) were issued on July 10, 2023 and took effect on August 15, 2023. They were issued jointly by seven agencies: the Cyberspace Administration of China (CAC), the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, the Ministry of Industry and Information Technology (MIIT), the Ministry of Public Security (MPS), and the National Radio and Television Administration (NRTA). The document runs 24 articles across five chapters: General Provisions, Technology Development and Governance, Service Specifications, Supervision and Inspection and Legal Liability, and Supplementary Provisions.

Scope. Article 2 covers organizations and individuals that use generative AI to provide public-facing content-generating services — text, images, audio, video — to the public within mainland China, and explicitly carves out in-house research and development that isn’t offered as a public service.
Overseas providers are in scope. Article 20 extends the measures to services provided from outside China to users inside China, giving regulators a basis to act against non-compliant offshore services.
Content obligations (Article 4). Generated content must uphold “socialist core values” and may not subvert state power, endanger national security, or promote separatism, terrorism, or extremism; providers must take measures against content that discriminates on ethnicity, belief, nationality, region, sex, age, occupation, or health; and providers must respect intellectual property and commercial ethics, and may not infringe others’ likeness, reputation, or privacy and personal-information rights.
Labeling, by cross-reference (Article 12). Providers must label generated images and video in line with the earlier Provisions on the Administration of Deep Synthesis Internet Information Services (effective January 10, 2023) — the Interim Measures don’t set their own labeling mechanics; they point at a separate, pre-existing regulation, which the 2025 Measures below now supersede in detail.
Algorithm filing and security assessment (Article 17). Services with “public opinion properties or social mobilization capacity” must complete a security assessment and file, register changes to, and deregister their algorithms with regulators — the same filing regime China runs for recommendation algorithms generally, extended here to generative models.

The 2025 Measures for Labeling of AI-Generated Synthetic Content

The Measures for Labeling of AI-Generated Synthetic Content (人工智能生成合成内容标识办法) were issued on March 7, 2025 and took effect on September 1, 2025 — a standalone, more detailed successor to the labeling mechanics the 2023 Interim Measures only cross-referenced. Four agencies issued it jointly: CAC, MIIT, MPS, and NRTA. It runs 14 articles and explicitly builds on the 2023 Deep Synthesis Provisions rather than replacing them outright.

Two label types, defined separately. An explicit label is added to the content or interface itself, presented as text, sound, or graphics, and “clearly perceivable by users.” An implicit label is embedded via technical means in the content file’s data, “not easily perceivable by users” — metadata, not a watermark you can see or hear.
Explicit labels are required across five content types (Article 4). Text gets a notification marker at the start, end, or within the content. Audio gets a voice prompt or rhythmic tone at the start, end, or an intermediate point. Images get a prominent marker at a suitable position. Video gets a clear marker on the opening frame plus a marker during playback. Virtual/generated scenes get a prominent marker on the initial interactive interface.
Implicit metadata is specified content, not open-ended (Article 5). The embedded file metadata must record generation-attribute information, the service provider’s name or code, a content identification number, and production-source information.
Tampering is separately prohibited. The measures bar maliciously deleting, altering, or forging the required labels — a distinct violation from simply omitting one.
Enforcement (Article 13) runs through existing law, not a new penalty schedule. The text assigns enforcement to cyberspace, telecom, public-security, and radio-and-television authorities “according to their duties” and directs them to “handle [violations] in accordance with relevant laws, administrative regulations, and departmental rules” — it does not itself set out a standalone fine table.

Where a national AI Law stands

China does not currently have a single, comprehensive national Artificial Intelligence Law comparable to the EU AI Act. We checked the 14th National People’s Congress’s own legislative-plan tracker (maintained by NPC Observer, most recently updated August 31, 2026) directly before publishing this guide: it lists no “Artificial Intelligence Law” entry in any of the NPC’s Category I, II, or III legislative projects, or among its unlisted projects. That means China’s AI governance to date has been built through administrative measures issued by the CAC and its co-regulators — the two instruments above, plus the 2023 Deep Synthesis Provisions and China’s broader cybersecurity and algorithm-recommendation rules — rather than through a statute passed by the national legislature. We’re stating this as a verified absence, not a prediction: if a national AI Law moves onto the NPC’s plan, that tracker is the place it would first show up, and this guide will be updated against it rather than against secondhand reporting.

How this compares to other jurisdictions

China’s approach is service- and content-specific rather than risk-tiered the way the EU AI Act’s high-risk compliance track is, and it carries real filing and assessment obligations rather than sitting at the voluntary end of the spectrum the way Singapore’s Model AI Governance Framework does. Its mandatory, two-tier (explicit-plus-implicit) content-labeling regime is also more prescriptive than Japan’s penalty-free AI Promotion Act, and it predates California’s SB 53 disclosure requirements by roughly two years on the services side and by several months on the labeling side. For the fuller jurisdiction-by-jurisdiction picture, see CASRAI’s AI regulations around the world guide; for how AI-content disclosure obligations look outside China specifically, see AI Disclosure Laws: The Legal Landscape vs. Publisher Policy.

The NIKOLAI angle: labeling as an unmapped Safeguard

NIKOLAI is CASRAI’s own frontier-AI-safety dictionary — an independent, unendorsed reference work, not an official record of any regulator’s or lab’s terminology, and certainly not an official record of China’s framework. Every NIKOLAI crosswalk row is a shadow mapping, CASRAI’s own interpretive reading of how a jurisdiction’s or organization’s language lines up with NIKOLAI’s elements, unless and until that organization files its own Mapping Declaration and it clears editorial review.

We checked directly before writing this: as of publication, neither NIKOLAI’s Safeguard element nor its Coverage level element — both in Track N6, Mitigations and security — carries a crosswalk row for China, the CAC, or any Chinese regulation. Safeguard is defined as “a technical or procedural measure intended to reduce misuse or misalignment risk, identified by type (e.g. safety training, filter, monitoring, access control), the risk domain it targets, and the deployment scope it applies to.” Mandatory content labeling fits that shape more directly than it fits any other NIKOLAI element we found: it is a named, typed technical-plus-procedural measure (explicit label plus implicit metadata), aimed at a specific risk domain (undisclosed synthetic content enabling fraud and disinformation), applied across a defined deployment scope (all public-facing text, image, audio, video, and virtual-scene generation in mainland China). What keeps this a candidate rather than an asserted mapping is that NIKOLAI’s existing Safeguard rows — Anthropic, OpenAI, Google DeepMind, xAI, Meta, the EU GPAI Code of Practice, California SB 53, and METR — are all classifier- or monitoring-style misuse safeguards a developer applies before or during generation, while China’s labeling requirement is a disclosure safeguard applied to the output itself. Whether NIKOLAI’s element definition is meant to stretch that far, or whether content-labeling regimes need their own element, is an open editorial question CASRAI hasn’t resolved, so we’re naming the candidate rather than filling in the row.

For how NIKOLAI’s shadow-mapping system works generally, see Comparing AI Safety Terms Across Frameworks: A NIKOLAI Crosswalk Guide.

Frequently asked questions

Does China require watermarks on all AI-generated content?

Yes, in the sense of a required marking system: since September 1, 2025, the Measures for Labeling of AI-Generated Synthetic Content require both a user-visible explicit label and embedded implicit metadata on AI-generated text, images, audio, video, and virtual scenes offered as public-facing services in China. “Watermark” isn’t the term the regulation itself uses, but the implicit-label mechanism functions the same way a metadata-based watermark does.

Did the 2025 labeling rules replace the 2023 Interim Measures?

No. They operate together. The 2023 Interim Measures govern how a generative-AI service has to operate overall (content standards, algorithm filing, security assessment) and originally pointed to the older 2023 Deep Synthesis Provisions for labeling mechanics. The 2025 Measures replace only that labeling cross-reference with a dedicated, more detailed regime; the Interim Measures’ other obligations are unaffected.

Does China have a comprehensive AI law like the EU AI Act?

Not currently. As verified against the 14th NPC’s own legislative-plan tracker (last updated August 31, 2026), no “Artificial Intelligence Law” appears among its Category I, II, or III projects. China’s AI governance runs through administrative measures issued by the CAC and co-regulators rather than through a single NPC statute.

Do foreign AI companies serving Chinese users have to comply?

The 2023 Interim Measures say yes explicitly: Article 20 extends the rules to services provided from outside China to users inside China. The 2025 labeling Measures are scoped the same way — to public-facing generative content services offered within China — regardless of where the provider is based.

What happens if a provider doesn’t comply with the labeling rules?

The 2025 Measures assign enforcement to China’s cyberspace, telecom, public-security, and broadcast regulators “according to their duties,” directing them to handle violations under existing laws, administrative regulations, and departmental rules, rather than setting out a standalone fine schedule inside the labeling Measures themselves.

Sources

Cyberspace Administration of China, notice on the Interim Measures for the Management of Generative AI Services (生成式人工智能服务管理暂行办法) — primary source for the issuance date (July 10, 2023), effective date (August 15, 2023), the seven issuing agencies, and the Article 4, 12, 17, and 20 obligations described above.
China Briefing, “How to Interpret China’s First Effort to Regulate Generative AI Measures” — used for corroboration of the Interim Measures’ scope and provider obligations.
Cyberspace Administration of China, notice on the Measures for Labeling of AI-Generated Synthetic Content (人工智能生成合成内容标识办法) — primary source for the issuance date (March 7, 2025), effective date (September 1, 2025), the four issuing agencies, the explicit/implicit label definitions, and the Article 4, 5, and 13 requirements described above.
NPC Observer, 14th National People’s Congress legislative-plan tracker — checked directly for the absence of any “Artificial Intelligence Law” entry, most recently updated August 31, 2026 as shown on the page at the time of this guide’s publication.
CASRAI’s own NIKOLAI element pages for Safeguard and Coverage level, checked directly for existing crosswalk rows before publication.

China’s AI Regulation: Interim Measures to the 2025 Labeling Rules