China's AI Content Labelling Rules Explained: Explicit Labels, Metadata and Platform Duties

What are China's AI content labelling rules?

China's AI content labelling rules are a set of binding measures and a mandatory technical standard that require AI-generated and synthetic content to carry visible and machine-readable labels. The core text is the Measures for Labeling AI-Generated Synthetic Content, issued jointly by the Cyberspace Administration of China (CAC), the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. The CAC published them on 14 March 2025, and they took effect on 1 September 2025.

The Measures cover text, images, audio, video, virtual scenes and other information produced with AI. They apply to service providers already regulated under China's rules on algorithmic recommendation, deep synthesis and generative AI, plus the platforms that distribute content and the app stores that list AI apps.

Alongside the Measures, the State Administration for Market Regulation and the Standardization Administration approved a mandatory national standard, GB 45438-2025, Cybersecurity technology: Labeling method for AI-generated synthetic content (announcement 2025 No. 3). It took effect on the same day. The Measures say what must be labelled and by whom; the standard says how. The National Information Security Standardization Technical Committee (TC260) also approved a practice guide on service provider coding rules for metadata labels on 14 March 2025, and adopted a further guide on implicit metadata labels for audio files on 28 August 2025.

What explicit labels must look like for text, audio, images, video and virtual scenes

An explicit label is one users can perceive: text, sound or a graphic placed in the content or in the interface around it. Article 4 of the Measures sets where it goes for each content type, and the standard adds the technical detail.

Labels must survive export. When a service lets users download, copy or export generated content, the file has to keep the explicit label. That clause matters for anyone receiving AI content outside the app that made it, because a downloaded image should still show its label.

Article 9 adds one exception. A provider may supply content without an explicit label, but only if the user agreement spells out the user's own labelling duties and responsibilities, and the provider keeps the related logs for at least six months.

Implicit labels: what goes into the file metadata

An implicit label is added to the file data by technical means and is not easily noticed by users. Article 5 requires providers to write it into the file metadata, meaning the descriptive information in the file header, and to include three things: attribute information about the content, the provider's name or code, and a content reference number. Providers are encouraged, though not required, to add digital watermarks as well.

TC260's coding-rules guide sets how the provider code is formed, which gives platforms a consistent way to read which service produced a file. Separate TC260 guides describe how the metadata label is written for particular file types, including the audio guide adopted in August 2025.

The design resembles a C2PA manifest in purpose: both put machine-readable origin information inside the file so software can read it later. They are different formats with different governance. C2PA is an industry specification with cryptographically signed manifests; China's implicit label is a government-mandated metadata field with a national provider code. A global platform may need to read both, and the guides to digital watermarking and metadata and provenance explain how each survives editing.

What platforms and app stores must do with AI content

Platforms that distribute content carry the heaviest operational duty. Under Article 6 they must check files for implicit labels and add a conspicuous notice to the content based on what they find. They must also write their own propagation metadata into the file, covering the content attributes, the platform's name or code and a content reference number, so the distribution chain is traceable.

Platforms must give users labelling tools and prompt them to declare AI content when they upload. Article 10 puts a matching duty on users: anyone publishing AI-generated content through a platform must declare it and use the platform's labelling function.

App stores have a narrower duty under Article 7. When reviewing an app for listing, they must ask whether it offers generative AI services and, if it does, check the provider's labelling materials. Providers must also explain their labelling methods in user agreements (Article 8) and submit labelling materials during algorithm filing and security assessment (Article 12).

How the 2025 rules build on the 2023 deep synthesis and generative AI measures

Labelling was already law in China before 2025. The deep synthesis provisions, in force since 10 January 2023, required providers to add identifiers that do not affect user experience to generated or edited content (Article 16). Article 17 required a conspicuous label wherever content could confuse or mislead the public, listing text that imitates a person, synthetic speech and voice imitation, face generation, face swapping and face or posture manipulation, and immersive realistic scenes. Article 18 banned using technical means to delete, alter or conceal those labels.

The Interim Measures for generative AI services, in force since 15 August 2023, covered generative services offered to the public in mainland China and, in Article 12, told providers to label generated images and video under the deep synthesis rules. Organisations developing generative AI without offering it to the public in mainland China were excluded.

The 2025 Measures turn those principles into an operating scheme: specific placement for each content type, a defined metadata payload, duties for platforms and app stores, and a mandatory standard to test against. Article 10 extends the old ban on removing labels, so no one may maliciously delete, tamper with, forge or conceal labels, or offer tools or services that help others do so. Article 13 leaves enforcement to the CAC, telecom, public security and broadcasting regulators under existing laws; the Measures themselves set no fines.

How China's labelling rules compare with the EU AI Act and India's rules

China, the EU and India now all require some form of visible disclosure plus machine-readable marking, but they assign the duties differently. China puts detailed duties on generators, distribution platforms and app stores, backed by one mandatory standard and a national provider code.

The EU AI Act Article 50 obligations apply from 2 August 2026. Article 50(2) requires providers of generative systems to mark outputs in a machine-readable format detectable as artificially generated, while Article 50(4) requires deployers to disclose deepfakes to people at first exposure. The European Commission says systems already on the market get until 2 December 2026 for the marking duty only.

India's amended IT intermediary rules (G.S.R. 120(E), in force from 20 February 2026) bring synthetically generated information into intermediary due diligence, with prominent labels and provenance metadata where technically feasible. The detailed page on India's synthetic content intermediary rules covers scope and open proposals.

Where labels stop working and forensic detection fits

Labels only help when they are present and intact. Metadata is lost when someone screenshots an image, re-encodes a video or uploads through a tool that discards headers. Content made with open-source models run locally, or by providers outside the scope of the rules, may never have been labelled. A missing label therefore tells a platform or reviewer nothing about whether the content is genuine.

The third tier in Article 6 relies on detection. A platform that labels content as suspected AI-generated needs some way to find generation traces when metadata is absent, and anyone handling disputes about that content needs evidence for the call. The same applies outside China: a marketplace reviewing a seller photo or an identity platform checking a selfie cannot assume an unlabelled file is authentic.

NeuroDetect is Neural Defend's forensic workspace for that review. An analyst uploads an authorized image, audio, video or document into a case, sees explainability heatmaps and attention views down to regions, frames and audio segments, and gets an LLM-assisted plain-language summary linked to the evidence, with the analyst making the final call. The case exports as a forensic report that records observations, uncertainty, provenance and model version, so a label decision can be explained later. For high-volume, real-time screening of face images and video, NeuroVerify returns a unified risk score from six signals in one API call.

Both run on ARCAS 1B, Neural Defend's own deepfake detection foundation model; the ARCAS paper reports pooled accuracy of 99.99% on GenImage and 97.65% on OpenFake. Detection gives evidence for a labelling or moderation decision. It does not satisfy the Measures by itself, and providers still have to add their own explicit and implicit labels.