Measures for Labelling AI-Generated Synthetic Content

CN·in-force·Grade A·checked 2026-10-11·source ↗

authority: Cyberspace Administration of China, Ministry of Industry and Information Technology, Ministry of Public Security and National Radio and Television Administration
published: 2025-03-07
effective: 2025-09-01
status: in-force

Adopted and signed by CAC, MIIT, MPS and NRTA on 7 March 2025, released online 14 March 2025 as the notice 国信办通字〔2025〕2号; effective 1 September 2025. China's most direct rule requiring AI-generated content to be labelled.

In one line

Text, images, audio, video and virtual scenes produced or synthesised by AI must carry labels: explicit labelling (visible to users) and implicit labelling (written into file metadata).

Who and what it governs

Object: AI-generated synthetic content — text, images, audio, video, virtual scenes.
Duty holders: service providers (the generating side and distribution platforms).

Key requirements

Add explicit labels so users can perceive that content is AI-generated.
Add implicit labels in file metadata so platforms and regulators can identify and trace it.
Providers must not delete, tamper with or forge labels, and must not help others evade labelling.

What the metadata has to carry (Articles 3, 5 and 6)

The elements are named by the measures themselves, not left to the standard:

Two kinds of label (Article 3) — an explicit label (显式标识) is added to the content or the interaction interface, presented as text, sound or graphics, and must be one "可以被用户明显感知到"; an implicit label (隐式标识) is inserted by technical means into the file data, where it is not readily perceptible to the user.
Production elements (Article 5) — acting under Article 16 of the Deep Synthesis Provisions, the provider writes the implicit label into the file metadata, which carries the content's attribute information, the provider's name or code, and a content identifier, together described as 制作要素信息. Digital watermarks are encouraged, not mandated. The same article defines file metadata as descriptive information embedded in the file header in a specific encoding, recording the file's origin, attributes and purpose.
Distribution elements (Article 6) — a provider that spreads network content must prompt the public wherever it verifies AI generation from the metadata, relies on a user's declaration, or detects generation traces; in those three cases it must also write attribute information, the platform's name or code, and a content identifier — 传播要素信息 — into the metadata. It must supply the labelling function and remind users to declare.
The standard attaches at Article 11 — labelling must additionally satisfy the requirements of the relevant laws, administrative regulations, department rules and mandatory national standards. That clause is what makes GB 45438-2025 binding rather than advisory.

Why it matters

This is where China moved fastest on "AI content must be identifiable": the measures took effect in September 2025, while the EU's comparable transparency duty only starts to apply in August 2026.

Companion technical standard

GB 45438-2025 《网络安全技术人工智能生成合成内容标识方法》 — a mandatory national standard (the portal classifies it 强标), published 2025-02-28, in force 2025-09-01, the same day as these measures; status 现行. The official record carries its own English title, Cybersecurity technology—Labeling method for content generated by artificial intelligence, and lists an English-language version; this card previously gave a translation of its own. Classifications CCS L80, ICS 35.030. Both the competent department and the technical committee (主管部门／归口部门) are 中央网络安全和信息化委员会办公室 — the cyberspace regulator owns the technical format — while the issuing bodies are 国家市场监督管理总局 and 国家标准化管理委员会.
- Official record: https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547
What the drafters say it does — the Q&A answers this directly in questions four and five: the measures state the management requirements and deliberately do not prescribe implementation, while the standard, issued and commenced together with them, "主要提出强制执行部分的标识具体实施方式和操作方法". Its content is given as three parts — how to add 文字、角标、语音、节奏 style explicit labels to text, images, audio, video and virtual scenes; how to write the metadata implicit label into a file; and a 预留字段 in that implicit label for protective information such as 标识完整性 and 内容一致性. The expert commentary adds that the standard fixes parameters such as an explicit label's duration and size and the elements an implicit label must contain. — B
The field names themselves are not in this repository. The portal does publish mandatory standards in full, but this one is served only as page images: the viewer pulls sprite sheets from a tile endpoint and reassembles each page in the browser. The tiles are reachable by script — the endpoint answers with WebP sheets once the reader's Referer is sent (re-checked 2026-10-12) — but what comes back is a picture, not text. The exact metadata keys, their encodings and the per-format parameters therefore remain unverified, and nothing in this card beyond the two bullets above is claimed about the standard's interior. ⚠

Relationship to other laws

Related: Interim Measures for the Administration of Generative AI Services (effective 2023-08-15), Provisions on the Administration of Deep Synthesis (effective 2023-01-10), Cybersecurity Law, Data Security Law.
Comparison: the EU AI Act's transparency obligations and the DSA's platform duties.

Official sources

Notice issuing the measures — 《关于印发〈人工智能生成合成内容标识办法〉的通知》, 国信办通字〔2025〕2号, dated 2025-03-07 and released 2025-03-14; the measures are its attachment and this page carries their full text: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
Announcement of the release: https://www.cac.gov.cn/2025-03/14/c_1743654685899683.htm
Q&A with the drafters: https://www.cac.gov.cn/2025-03/14/c_1743654685896173.htm
Expert explainer Putting a digital label on AI-generated content: https://www.cac.gov.cn/2025-09/05/c_1758792061408012.htm
Standard record, GB 45438-2025 (国家标准全文公开系统): https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547

Source quotes

Every sentence below is copied from the official text and pinned to the snapshot in snapshots/ that this card was written against. node scripts/check-quotes.mjs re-checks them offline, so a quote that stops matching the source fails a check instead of quietly becoming the record. The prose above is a rendering; these are the anchors.

隐式标识包含生成合成内容属性信息、服务提供者名称或者编码、内容编号等制作要素信息 — Article 5 — the production elements (snapshot: snapshots/cn/ai-generated-content-labeling-measures.txt)
文件元数据是指按照特定编码格式嵌入到文件头部的描述性信息 — Article 5 — what file metadata is (snapshot: snapshots/cn/ai-generated-content-labeling-measures.txt)
鼓励服务提供者在生成合成内容中添加数字水印等形式的隐式标识 — watermarks encouraged, not required (snapshot: snapshots/cn/ai-generated-content-labeling-measures.txt)
应当在文件元数据中添加生成合成内容属性信息、传播平台名称或者编码、内容编号等传播要素信息 — Article 6 — the distribution elements (snapshot: snapshots/cn/ai-generated-content-labeling-measures.txt)
还应当符合相关法律、行政法规、部门规章和强制性国家标准的要求 — Article 11 — where the mandatory standard attaches (snapshot: snapshots/cn/ai-generated-content-labeling-measures.txt)
本办法自2025年9月1日起施行 — Article 14 (snapshot: snapshots/cn/ai-generated-content-labeling-measures.txt)
Cybersecurity technology—Labeling method for content generated by artificial intelligence — the standard’s own English title (snapshot: snapshots/cn/gb-45438-2025-record.txt)
中央网络安全和信息化委员会办公室 — competent and归口 department (snapshot: snapshots/cn/gb-45438-2025-record.txt)
国家市场监督管理总局、国家标准化管理委员会 — issuing bodies (snapshot: snapshots/cn/gb-45438-2025-record.txt)
主要提出强制执行部分的标识具体实施方式和操作方法 — what the standard does, in the drafters' own words (snapshot: snapshots/cn/ai-generated-content-labeling-measures-qa.txt)
两者同步推出，于2025年9月1日同步实施 — measures and standard issued and commenced together (snapshot: snapshots/cn/ai-generated-content-labeling-measures-qa.txt)
添加文字、角标、语音、节奏等显式标识的方法 — explicit-label methods per content type (snapshot: snapshots/cn/ai-generated-content-labeling-measures-qa.txt)
在元数据隐式标识设置了预留字段，可用于记录标识完整性、内容一致性等安全防护信息 — the reserved metadata field (snapshot: snapshots/cn/ai-generated-content-labeling-measures-qa.txt)
（如显式标识的时长、尺寸，隐式标识的包含要素等） — what the standard fixes, per the expert commentary (snapshot: snapshots/cn/ai-generated-content-labeling-measures-expert.txt)
2025-09-01 — in force date (snapshot: snapshots/cn/gb-45438-2025-record.txt)

Unverified / TODO

Field-level specification in GB 45438-2025 — the exact implicit-label metadata keys, their encodings, and per-format parameters such as an explicit label's duration and size — is not transcribed. The standard's full text is served only as page images: openstd's viewer fetches sprite sheets from a tile endpoint and reassembles each page in the browser from 10×10 tiles. Re-checked 2026-10-12, and the correction matters — the tiles are reachable by script (the endpoint answers with WebP sheets once the reader's Referer is sent, and the tiles can be cut out by their background-position and reassembled into page images). What cannot be had is text: the reassembled page is a picture, so no snapshot of it exists for check-quotes.mjs to compare against, and any transcription from it — by script or by eye — is unverifiable and therefore inadmissible here. ⚠ [gap:not_transcribable]
The supporting practice guides (metadata text requirements, detection methods) have not been located against an official source. ⚠ [gap:not_located]
Confirm the specific penalty provisions and amounts — resolved: neither the Generative AI Interim Measures (Article 21) nor the Deep Synthesis Provisions (Article 22) sets a fine amount of its own; both refer punishment to the general laws, with suspension as the operational sanction