China rolls out mandatory labels for AI-generated content

On September 1, 2025, rules identifying AI-generated or AI-altered content took effect in China. The original source supports the documentary core of the event; a labeling duty does not guarantee that every platform will detect or preserve provenance correctly.

Major platforms such as WeChat, Douyin — the Chinese version of TikTok — and Weibo have rolled out changes to comply with the rules, South China Morning Post reports. The measure applies both to companies offering AI tools and to the services distributing their output, as well as to the users who publish it.

Two labels for the same piece of content

The measures were published in March by the Cyberspace Administration of China (CAC), together with the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. They take effect alongside a national technical standard spelling out how the markers must be applied.

The system distinguishes between two types of identification:

• The explicit label must be perceptible to the user. It could be a notice alongside a text, a graphic marker over an image, or a message added to a video or audio file.
• The implicit label is added to the file’s metadata — the technical information that accompanies it even though it is not normally visible. It may include details about its origin, the provider used and the type of synthetic content.

The combination is intended to address a common weakness of visible watermarks: they can be cropped, erased or covered up. Metadata allows for automated verification, although it is not foolproof either, as it can be lost when a file is copied, compressed or transformed.

The rules prohibit removing, tampering with or falsifying these identifiers. They also bar companies from deliberately providing tools designed to strip them.

Platforms must verify and complete the labelling

Responsibility does not end with the AI generator. When a platform finds the relevant identification in a file’s metadata, it must display a notice visible to the public. If it detects signs that content is synthetic despite the absence of a technical marker, it may also label it as suspicious or likely AI-generated.

Users, for their part, must disclose their use of artificial intelligence when publishing content. WeChat has said its creators will have to voluntarily flag such content; for posts that have not been identified, the platform will display warnings so readers can “exercise their own judgment” about their authenticity.

This division of responsibilities is one of the most significant aspects of China’s model. It does not rely solely on creators to act properly: it builds controls into AI providers, social networks and app distribution channels. App stores may also require documentation demonstrating that generative services comply with the rules.

For companies, this means changes throughout the technical chain. It is not enough to place a sentence beneath an image: identifiers must be preserved during the file’s generation, export, upload and redistribution.

A response to AI-generated fraud and disinformation

Beijing presents the rules as a tool against disinformation, scams, impersonation and intellectual property violations. Chinese regulators believe that deepfakes — manipulated recordings designed to imitate a person’s face or voice — can threaten both individual and national security.

Labelling can provide context, but it does not determine whether a post is true. AI-generated content can be accurate, while an authentic photograph can be used to deceive. Its effectiveness will also depend on platforms recognising markers from other services and preventing them from disappearing during editing or downloading.

The regulation also expands the Chinese digital ecosystem’s capacity for oversight. The same mechanisms that make it possible to trace the origin of a fraudulent video also make it easier to attribute and control the circulation of content. In China, transparency about AI use is thus folded into a broader policy of internet surveillance and moderation, including the CAC’s annual Qinglang campaign.

China moves ahead of Europe’s timetable

The European Union will also require certain AI-generated or manipulated content to be identifiable in a machine-readable format. Article 50 of the AI Act also includes disclosure obligations for deepfakes and certain texts of public interest, but those provisions will not generally apply until August 2026.

China is moving nearly a year ahead on implementation and is adopting an operational approach from today that covers generators, platforms and users. That does not mean the two systems are equivalent: the European framework distinguishes obligations according to who creates or publishes the content and includes exceptions, while China’s system is integrated into a far more centralised internet-control regime.

The decisive test begins now. Regulators will have to show that markers survive as files move between applications and that platforms apply consistent standards to suspicious content. For WeChat, Douyin, Weibo and AI providers, labelling is no longer a voluntary feature; it is now part of the mandatory infrastructure for publishing content.

Turning the headline into a check

Policy reading begins with authority. Separate statute, regulation, guidance, plan, contract and a platform's private setting. Then identify who is bound, from which date and before which body. a labeling duty does not guarantee that every platform will detect or preserve provenance correctly. Without that classification, a policy priority can be misreported as an enforceable right or prohibition.

The text also needs an implementation chain: published rule, technical specification, budget, owner and evidence of compliance. The existence of a duty does not show that every tool can perform it or that enforcement detects every breach. To assess how to check who labels, which metadata travels and what happens when the file is transformed, look for the points where the chain can break.

Exceptions and transitions are part of the rule. Older products, smaller actors, research, security or market dates may receive different treatment. A practical check preserves the applicable article or section and explains why the case belongs there. A list of duties without subject or timeline can prompt the wrong action.

What the record must preserve

The final indicator is observable conduct: a published document, enabled option, submitted report, awarded contract or appealable sanction. Intermediate announcements are dated, but their effect is not brought forward. This lets the reader follow change without confusing intent, technical capacity and effective compliance.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to check who labels, which metadata travels and what happens when the file is transformed. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.