China's Labelling, Filing and Content Rules

China regulates AI through a series of targeted administrative measures rather than a single statute: algorithmic recommendation (2022), deep synthesis (2023), generative AI services (2023) and content labelling (2025). Each was issued by the Cyberspace Administration of China with other ministries, and each applies to services offered to the public within mainland China.

The four instruments

The pattern is worth understanding because it predicts what comes next. Each measure is narrow, issued by administrative bodies rather than by the legislature, effective within months, and framed as management of an information service rather than as regulation of a technology. The regulated object is the service offered to the public, not the model.

There is no comprehensive AI law

A draft Artificial Intelligence Law has appeared on State Council legislative work plans since 2023, and scholarly draft texts have circulated. As at this review date, no comprehensive AI statute has been enacted in China. What binds is the administrative measures above, the Cybersecurity Law, the Data Security Law and the Personal Information Protection Law.

There is also a national AI safety governance framework published by the national technical committee on cybersecurity standardisation, first issued in 2024 and revised since. It is a framework document rather than a binding rule, and it is useful mainly as a statement of how the risk taxonomy is officially conceived.

The filing regime, which gates the launch

This is the part that determines whether a service can go live, and it is where compliance in China differs most sharply from the EU or US models. There is no conformity assessment and no CE mark; there is a registration you must complete before offering the service to the public.

1. Determine whether the service has public opinion attributes or social mobilisation capacity. In practice, any public-facing service that generates content for users is treated as having them. Internal enterprise tools generally are not.
2. Conduct a security assessment under the existing regime for internet information services with those attributes, and file the results.
3. File the algorithm through the CAC’s algorithm filing system, providing the provider’s details, the algorithm’s type, its application scenario, a self-assessment report and content to be published.
4. Complete the large model filing for public-facing generative AI services. The CAC publishes lists of registered generative AI services in batches; appearing on such a list is the public signal that a service has completed the process.
5. Maintain and update. Material changes to the algorithm or the service require a further filing, and the filing number must be displayed on the service.

The commercial consequence: the timeline to launch a public generative AI service in China is set by this process, not by engineering, and it is a genuine barrier for foreign entrants. Serving Chinese users from outside the mainland does not avoid it in principle and creates separate cross-border data issues.

Content obligations

The generative AI measures impose substantive obligations on the content a service produces and on the data used to train it. In summary, without citing article numbers:

• Content standards. Generated content must uphold core socialist values and must not incite subversion of state power, endanger national security or the public interest, promote terrorism, extremism, ethnic hatred or discrimination, violence, obscenity, or false and harmful information.
• Non-discrimination. Measures must be taken during algorithm design, training data selection, model generation and optimisation to prevent discrimination on grounds including ethnicity, belief, nationality, region, gender, age and occupation.
• Training data lawfulness. Data and foundation models used must come from lawful sources; intellectual property rights must not be infringed; personal information requires consent or another lawful basis; and measures must be taken to improve the quality, authenticity, accuracy, objectivity and diversity of training data.
• Provider responsibility. The service provider bears responsibility as the producer of the content and as the handler of personal information, must have a complaints mechanism, must stop generating and take remedial action on discovering unlawful content, and must report to the authorities.
• User protections. No excessive collection of personal information, no unlawful retention of input data that identifies users, and no unlawful provision of user inputs and usage records to others.

The labelling measures

The labelling regime, in force since 1 September 2025, is the most concrete and most exportable of the Chinese rules, and it goes further than the EU’s Article 50 transparency duties in two ways.

• Two kinds of label are required, not one. An explicit label, perceptible to the user — text on an image, a spoken or written notice at the start of audio or video, a marker in a chat interface — and an implicit label embedded in the file metadata recording that the content was AI-generated, along with the provider and a content identifier.
• The duty runs down the distribution chain. Platforms that distribute content must check for implicit labels, and where content is or may be AI-generated, mark it as such for users. Application stores are expected to verify that the applications they list comply.

Users may request content without an explicit label for legitimate purposes, but the provider must retain a record and the user takes on responsibility. Removing, tampering with, forging or concealing labels is prohibited, as is providing tools whose purpose is to do so.

If you build a product that handles generated media, the design consequence is that provenance metadata has to survive your pipeline. A system that strips EXIF on upload, or re-encodes video without carrying metadata forward, breaks the implicit label and is a compliance problem in this jurisdiction even though it is nobody’s problem elsewhere.

Data protection and cross-border transfer

The Personal Information Protection Law, in force since 1 November 2021, is the general data law and its automated decision-making provision (Article 24) is the analogue of GDPR Article 22: automated decision-making must be transparent and the results fair and reasonable, differential treatment in transaction terms is prohibited, personalised push must offer an option not based on personal characteristics or a way to refuse, and an individual may require an explanation and refuse decisions made solely by automated means where the decision has a major effect on their rights.

Cross-border transfer of personal data out of mainland China runs through a separate regime — security assessment, standard contract or certification, depending on volume and sensitivity — which was relaxed in 2024 with exemptions for certain volumes and scenarios. If your architecture routes Chinese user prompts to a model hosted outside the mainland, this is the analysis that governs it and it is not a formality.

What this means for a foreign company

Three practical conclusions.

If you do not offer services to the public in mainland China, none of this binds you. The measures apply to services provided to the public within the mainland. A product with incidental Chinese-speaking users elsewhere is not caught by that alone.

If you do, the filing regime is the gate, and it is usually done through a local entity or partner. There is no route that treats the launch as a purely technical event.

If you build media tooling anywhere, the labelling measures are worth engineering to regardless, because the direction of travel is common to China, the EU’s Article 50 and California’s provenance legislation, and because implicit-label preservation is much cheaper to design in than to retrofit. See the status of the other major jurisdictions for how that convergence looks elsewhere.