Last reviewed
Jun 7, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 1,842 words
Add missing citations, update stale details, or suggest a clearer explanation.
China's AI-generated content labeling rules are a mandatory regime that requires synthetic media produced by generative AI services to carry both visible and machine-readable markings before it reaches the public. The core instrument is the Measures for the Identification of AI-Generated Synthetic Content (人工智能生成合成内容标识办法), issued on 14 March 2025 by the Cyberspace Administration of China (CAC) together with three other regulators, and effective 1 September 2025. The Measures are paired with a mandatory national technical standard, GB 45438-2025, which specifies exactly how the labels must be encoded. The framework is widely described by international law firms as one of the most comprehensive synthetic-media labeling regimes in force anywhere, predating the equivalent transparency obligations of the EU AI Act by roughly a year. It distinguishes two label types, explicit (visible or audible) and implicit (embedded in file metadata or as a digital watermark), and assigns duties to AI service providers, content-distribution platforms, app stores, and ordinary users.
The labeling Measures sit on top of an existing layer of Chinese internet rules rather than replacing them. Two earlier instruments matter most. The Provisions on the Administration of Deep Synthesis Internet Information Services (the "Deep Synthesis Provisions"), issued 25 November 2022 and effective 10 January 2023, first required providers of services that generate or significantly alter content to add prominent markings so that the public is aware the material is synthesized. The Interim Measures for the Management of Generative AI Services (the Generative AI Interim Measures), effective 15 August 2023, then required generative-AI providers to label images, video, and other outputs in line with the Deep Synthesis Provisions. Both measures gestured at labeling but left the technical specifics open. The 2025 Measures and GB 45438-2025 fill that gap by turning the general obligation into a concrete, testable standard. The whole stack is anchored in higher-level statutes including the Cybersecurity Law and forms part of the broader system of Chinese AI regulation. China had also signaled the direction earlier: an initial draft of the labeling rules was released for public comment in September 2024 before the finalized version appeared in March 2025.
The Measures were jointly promulgated by four bodies: the CAC, the Ministry of Industry and Information Technology (MIIT), the Ministry of Public Security (MPS), and the National Radio and Television Administration (NRTA). They apply to "AI-generated synthetic content," defined broadly to cover text, images, audio, video, and virtual scenes produced or synthesized using AI.
GB 45438-2025, titled "Cybersecurity technology - Labeling method for content generated by artificial intelligence," is the accompanying mandatory national standard. According to the China Standards database, it was approved and issued by the State Administration for Market Regulation (SAMR) and the Standardization Administration of China (SAC) on 28 February 2025, released alongside the Measures on 14 March 2025, and made effective on the same date as the Measures, 1 September 2025. A companion document, a "Cybersecurity Standard Practice Guideline" on encoding rules for service providers, was published to give implementation detail. As a mandatory ("GB", as opposed to recommended "GB/T") standard, GB 45438-2025 is itself legally binding, not merely advisory.
The regime's defining feature is its two-track labeling model. Explicit labels are perceptible to a human reader or viewer. Implicit labels are written into the file so that machines and platforms can detect AI provenance even when no visible mark survives.
| Aspect | Explicit label | Implicit label |
|---|---|---|
| Nature | Visible or audible marking a person can perceive | Data embedded in file metadata, or a digital watermark |
| Where it appears | On screen or in the audio: text, graphic, superscript "AI", or spoken/Morse cue | Inside the file's metadata fields (and optionally as a watermark in the content itself) |
| Content types | Text, images, audio, video, virtual scenes, interactive interfaces | All generated/synthesized files |
| When required | Mandatory where content could mislead or confuse the public | Mandatory for files; must persist through download, copy, and export |
| Set by | Generation service providers (and users on declaration) | Generation and distribution service providers |
Under GB 45438-2025, explicit text markings must include "artificial intelligence" or "AI" together with a "generation" or "synthesis" element, placed at the start, middle, or end of the text in a clearly distinguishable font and color. For images and video, a text prompt is required at an edge or corner, with the standard specifying that label text height should be no less than 5 percent of the image's shortest side and that a video label should display for at least 2 seconds at normal playback speed. Audio can be labeled with a spoken cue at a normal speaking rate or with a Morse-code rhythm representing "AI" (short long, short short short). For conversational services such as chatbots, a visible label must sit at a suitable point in the generated text, and any downloadable file must carry the explicit label as well.
Implicit labels are encoded in file metadata using a standardized JSON structure under a field named "AIGC". The required elements include a "Label" certainty flag, a "ContentProducer" identifying the generation service provider, a "ProduceID" content number, and, when the file is later distributed, a "ContentPropagator" and "PropagateID" for the distribution platform. The certainty flag encodes a three-tier classification: confirmed AI-generated content, possibly AI-generated, and suspected. The standard also permits implicit marking via digital watermark embedded in the content, though it does not mandate a single watermarking algorithm. This metadata-plus-watermark approach is what makes the regime auditable: platforms can read provenance programmatically rather than relying only on a visible badge that a re-upload might strip.
The Measures distribute duties across four categories of actor.
| Actor | Core obligation |
|---|---|
| Generation service providers | Add explicit labels (where the content may mislead) and embed implicit metadata labels in every generated file; keep implicit labels intact through download, copy, and export |
| Content distribution / propagation platforms | Verify whether content carries labels, surface a notice to users for material that is or appears to be AI-generated, and add labels to detected-but-unlabeled content; supply tools for users to declare AI content |
| App distribution platforms (app stores) | During app review, require providers to state whether they offer generative-AI services and check their content-labeling materials before listing |
| Users | Proactively declare AI-generated content they post and use the platform-provided labeling tools |
A key prohibition runs through the regime: no one may maliciously delete, tamper with, forge, or conceal the required labels, nor provide tools to do so. This anti-tampering rule is what platform implementations leaned on most heavily in practice.
The Measures and GB 45438-2025 both took effect on 1 September 2025. China's largest platforms moved to comply on or before that date. According to reporting collected by SiliconANGLE and others, WeChat (which Tencent cites at roughly 1.4 billion monthly active users) deployed automated detection for unlabeled AI content and reiterated the ban on deleting, tampering with, forging, or concealing labels; Douyin, the domestic version of TikTok, added creator labeling tools and metadata-based detection to flag untagged AI video; and Weibo introduced both a self-tagging tool and a button for users to report unlabeled AI content posted by others. Xiaohongshu (RedNote), Bilibili, Zhihu, and e-commerce platforms such as Tmall and JD.com were also within scope, and Tencent's Yuanbao chatbot applied explicit and implicit tags to the text, images, and video it produces. Enforcement flows through the existing internet-governance toolkit rather than a bespoke penalty schedule in the Measures themselves: non-compliant content can be removed, and providers can face regulatory investigation, business suspension, and, for repeat or serious violations, action against their operating licenses, with the underlying Cybersecurity Law and sectoral rules supplying the sanction basis.
The Chinese regime is notable for combining a binding administrative measure with a binding technical standard, so that "label your AI content" comes with an enforceable specification of how. That pairing is what leads many analysts to call it the most comprehensive synthetic-media labeling framework currently operative. The closest comparison is the EU AI Act. Article 50 of the AI Act requires providers of generative systems to mark outputs in a machine-readable format detectable as artificially generated, and requires deployers to disclose deepfake image, audio, and video; those obligations apply from 2 August 2026, and the European Commission published a first draft Code of Practice on transparency in December 2025 to guide compliance. In broad strokes, both regimes pursue the same transparency goal through a mix of visible disclosure and machine-readable marking. The Chinese version differs in being already in force, in prescribing concrete formatting through GB 45438-2025, and in placing explicit duties on distribution platforms and app stores rather than concentrating obligations on model providers and deployers. For multinational AI companies, the practical effect is that any generative service touching the Chinese market must implement both a visible label and a standardized metadata payload, and must keep that payload intact across the content's lifecycle, a requirement that has pushed provenance-by-metadata toward becoming a baseline expectation in AI regulation globally.