Measures for the Identification of AI-Generated and Synthesized Content
📑 Legal hierarchy: Level 3 · Departmental rule | Issuance: CAC, MIIT, MPS, NRTA (four-ministry joint issuance) | Effective: 2025-09-01 | Character: hard law
⚠️ Hierarchy note: This instrument is a departmental rule, jointly issued by four ministries. It is neither a NPC law nor a State Council administrative regulation. Its penalties land via reference to upstream laws such as the Cybersecurity Law. See Index of Chinese Rules.
English Summary
Section titled “English Summary”The Measures for the Identification of Artificial Intelligence Generated and Synthesized Content (《人工智能生成合成内容标识办法》), jointly issued by the Cyberspace Administration of China (CAC) and three other ministries on 2025-03-07 and effective 2025-09-01, impose a dual-track labeling regime on AI-generated content: explicit labels (human-readable watermarks or text prompts) and implicit labels (machine-readable metadata embedded in the file). Obligated parties include (i) generative AI service providers, (ii) platforms that host and distribute such content, and (iii) end-users uploading synthesized content. The Measures are operationalized by the mandatory national standard GB 45438-2025, which defines the technical fields for implicit labels. Enforcement follows the penalty framework of the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law.
Overview
Section titled “Overview”The Labeling Measures are the first systematic departmental rule on the duty to label AI-generated and synthesized content in China, resolving the question of how the earlier principled requirements to “label such content” in the Deep Synthesis Provisions and the Generative AI Interim Measures are to be operationalized.
Three breakthroughs:
- From “shall label” to “dual-track labeling”: for the first time, explicit labels (human-perceptible) and implicit labels (machine-readable) are both set as duties. The former serves public notice; the latter enables automated detection across platforms and through content flows.
- From single-point provider duty to chain-wide coverage: duty subjects are extended from “service providers” to service providers + distribution platforms + uploading users. Distribution platforms carry a verification duty and, where labels are missing, shall proactively add them.
- Bundling with a technical standard upstream: effective the same day as GB 45438-2025, which specifies the implicit-label fields (including provider name, content ID, generation date), enabling interoperability across providers.
- Territory: services within China and offshore services aimed at domestic users (actual enforcement intensity remains to be seen).
- Activities: four types of synthesized content — text, image, audio, video.
- Subjects:
- generative AI service providers (including deployers of large models);
- network platforms that host and distribute user-generated content;
- users who upload synthesized content to platforms.
Core Duties
Section titled “Core Duties”On service providers
Section titled “On service providers”- Explicit labels: add an “AI-generated” prompt or equivalent at a conspicuous location in the generated content (text label, corner mark, or watermark).
- Implicit labels: embed fields in the file metadata or within the content itself: provider name, content ID, generation date, and other fields, complying with GB 45438-2025.
- No anti-removal design: shall not deliberately make labels trivially removable or destructible.
On distribution platforms
Section titled “On distribution platforms”- Verification: check, upon receipt of uploaded content, whether compliant implicit labels are present.
- Proactive tagging: where labels are missing, proactively add a notice such as “the platform declares this content may be synthesized.”
- No bad-faith interference: shall not delete, tamper with, or hide another party’s compliant labels.
On uploading users
Section titled “On uploading users”- Proactively declare content to be AI-synthesized.
- Shall not maliciously remove labels applied by platforms or service providers.
Enforcement and Penalties
Section titled “Enforcement and Penalties”Basis:
- Cybersecurity Law Article 68 (penalties for information service violations);
- Data Security Law;
- Personal Information Protection Law.
Penalty gradient (from the above laws): warning → order to rectify → fine → suspension of operations → revocation of permits. Serious cases engage public-security administrative penalties or criminal liability.
Competent authorities: CAC leads; MIIT (compute and infrastructure), MPS (criminal), and NRTA (audiovisual) coordinate within their mandates.
Relationship with Other Rules
Section titled “Relationship with Other Rules”- Deep Synthesis Provisions (2023): the upstream principled rule; these Measures elaborate its Article 17 “shall apply conspicuous labeling.”
- Generative AI Interim Measures (2023): a sister rule on provider-side filing and security assessment; these Measures focus on the output layer (labeling), while the Interim Measures focus on the service layer (compliance).
- GB 45438-2025: the mandatory national standard defining the implicit-label fields. This site classifies it as soft law (see reasoning) while flagging its mandatory character.
- Personal Information Protection Law: where synthesized content involves faces, voices, or other sensitive information, PIPL Article 28 applies jointly.
Debates and Commentary
Section titled “Debates and Commentary”- Offshore applicability: how will these Measures be enforced against offshore providers serving users within China? The alignment with the mainland TikTok counterpart remains unclear.
- Robustness of implicit labels: will GB 45438’s implicit labels survive screenshots, compression, or re-editing? Scholars have raised doubts about technical feasibility.
- Artistic and satirical carve-outs: unlike EU AI Act Article 50, these Measures do not expressly exempt artistic or satirical works. How are the boundaries to be drawn in practice?
- Bounds of platform verification responsibility: requiring platforms to verify implicit labels on every upload is technically expensive; the feasibility for small and mid-sized platforms remains to be observed.
Source Text and Translations
Section titled “Source Text and Translations”This site does not produce its own full translations. The Chinese text of record is the official release; for English, we rely on academically respected secondary translations. For citation of specific articles, please return to the primary and authoritative sources below.
| Language | Source | Link |
|---|---|---|
| Chinese (original) | CAC | cac.gov.cn |
| Chinese (archived copy) | This site | biaozhi-banfa-2025-03-07.html |
| English | China Law Translate (Jeremy Daum, Paul Tsai China Center, Yale) | chinalawtranslate.com/en/ai-labeling |
| English (structured) | Regulations.AI | regulations.ai/…/MIASCXX-2025 |
| Companion national standard | National Public Service Platform for Standards Information | GB 45438-2025 |
Version History
Section titled “Version History”| Date | Event |
|---|---|
| 2024-09 | Draft for public comment released |
| 2025-03-07 | Official release |
| 2025-09-01 | Effective |
| 2026-04-21 | First archived on this site |
Cited Sources
Section titled “Cited Sources”Primary
Section titled “Primary”- CAC release: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
- GB 45438-2025 Cybersecurity Technology — Labeling Method for AI-Generated and Synthesized Content (mandatory national standard): https://openstd.samr.gov.cn/bzgk/gb/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547&refer=outter
- CAC press Q&A (released same day, 2025-03-07): https://www.cac.gov.cn/2025-03/14/c_1743654735022812.htm
Authoritative English translations
Section titled “Authoritative English translations”- China Law Translate (Jeremy Daum, Paul Tsai China Center, Yale Law School): https://www.chinalawtranslate.com/en/ai-labeling/
- Regulations.AI (structured version): https://regulations.ai/regulations/RAI-CN-NA-MIASCXX-2025
Major international commentary
Section titled “Major international commentary”- Covington & Burling Inside Privacy: https://www.insideprivacy.com/international/china/china-releases-new-labeling-requirements-for-ai-generated-content/
- DGC Briefings: https://dgcbriefings.substack.com/p/china-finalised-measures-for-labelling
- Technology’s Legal Edge (Perkins Coie): https://www.technologyslegaledge.com/2025/03/china-released-new-measures-for-labelling-ai-generated-and-synthetic-content/
Related research
Section titled “Related research”- Zhang Linghan 张凌寒 (China University of Political Science and Law / CUPL; Renmin University) — on the hierarchical relationships among the Labeling Measures, Deep Synthesis Provisions, and Interim Measures, and jurisprudential analysis of the “explicit + implicit dual track” as an institutional innovation.
- Zhu Yue 朱悦 and Dai Xin 戴昕 (CUPL / Peking University) — institutional challenges of implicit labeling and machine-readable metadata under cross-border interoperability.
- Matt Sheehan (Carnegie Endowment) — institutional-pathway analysis in the China’s AI Regulations and How They Get Made series.
- Paul Triolo (DigiChina) — comparative institutional analysis of China’s mandatory dual-track labeling against US and EU regimes.
- CAIDP (Center for AI and Digital Policy) — lists the Chinese Labeling Measures as a reference example of “mandatory content provenance” in the AI and Democratic Values Index 2025.
Cite this page (generated 2026-04-21):
Comparative AI. Commentary on the Measures for the Identification of AI-Generated and Synthesized Content. Accessed YYYY-MM-DD. https://comparativeai.org/rules/china/biaozhi-banfa/