China's AI Regulation: What Foreign Teams Operating in China Need to Know

The Generative AI Measures, deep-synthesis and algorithm rules, and the 2025 labelling standard, plus what actually applies to non-Chinese providers.

Educational content, not legal advice. AI regulation is fast-moving and jurisdiction-specific. Verify every obligation and deadline against the primary legal text and consult qualified counsel before making compliance decisions. Accurate as of the research date shown in the metadata below.

A different model: vertical rules, not one AI Act

China does not regulate AI through a single horizontal statute like the EU. Instead it has built a stack of targeted, iterative rules issued mainly by the Cyberspace Administration of China (CAC) alongside other ministries, each aimed at a specific technology or risk. A unified national AI law has been discussed and studied, but as of mid-2026 the operative regime is this set of vertical measures. Understanding them individually matters more than waiting for an omnibus.

The Interim Measures for Generative AI Services

Effective 15 August 2023 and issued jointly by the CAC and six other regulators, the Interim Measures are the anchor rule for generative AI. They apply to the provision of generative AI services to the public within mainland China. Core obligations include:

• Content standards: generated content must uphold core socialist values and must not endanger national security, incite subversion, or spread prohibited content; providers must moderate accordingly.
• Lawful training data: use data and foundation models from legitimate sources, respect intellectual property and personal-information rights, and improve data quality and accuracy.
• Security assessment and algorithm filing: services with 'public-opinion attributes or social-mobilisation capabilities' must complete a security self-assessment and file their algorithm with the CAC before launch.
• Labelling: label AI-generated content in line with the deep-synthesis rules and the 2025 labelling standard.
• User protection: verify aspects of user identity where required, protect minors, handle personal information under China's PIPL, and provide complaint mechanisms.

A notable feature is the graduated, 'inclusive and prudent' tone: the final Measures softened some draft provisions and carved out research and internal (non-public-facing) use, signalling an intent to regulate without smothering domestic industry.

Deep synthesis and algorithm registry rules

The Deep Synthesis Provisions (effective January 2023) govern technologies that generate or edit synthetic media, text, images, audio, video and virtual scenes. They require conspicuous labelling of content that could confuse the public, consent for editing biometric information such as faces and voices, and record-keeping. The Algorithm Recommendation Provisions (effective March 2022) created China's algorithm filing registry: providers of recommendation algorithms with public-opinion or social-mobilisation characteristics must register details of their algorithms with the CAC. That registry has become the backbone of Chinese AI oversight, and filing is a recurring precondition across the later rules.

The 2025 labelling standard

On 14 March 2025 the CAC released the Measures for Labelling AI-Generated Content, paired with the mandatory national standard GB 45438-2025. Both took effect on 1 September 2025. The regime requires a dual approach:

• Explicit labels: human-visible notices (for example 'AI-generated') placed at the start, end or an appropriate position in text, images, audio and video.
• Implicit labels: machine-readable markers embedded in file metadata, typically identifying the service provider and a content reference number, with watermarking where feasible.

Crucially, the duty is shared: not only the generating service but also distribution platforms must verify and surface labels for AI content, and remove or sanction non-compliant content. This mirrors, and in some respects predates and exceeds, the labelling ambitions of the EU AI Act's Article 50.

What foreign teams must actually watch

Access reality: many foreign generative AI products are not authorised for public provision inside mainland China. If your service is reachable by the Chinese public, these rules can apply to you; if it is blocked or geofenced, your exposure is different. Map where your product is actually offered before assuming either way.

1. Scope by audience, not incorporation. The Interim Measures target services provided to the public in mainland China. A foreign provider serving Chinese users can be in scope; internal enterprise or R&D use is generally treated differently.
2. Filing and assessment are gatekeepers. Public-facing generative services with public-opinion or social-mobilisation reach need CAC security assessment and algorithm filing before launch, which is a material lead-time item, not a formality.
3. Content moderation is a hard requirement, not a nice-to-have. The content standards are enforced, and they differ substantially from Western norms; a compliant deployment needs China-specific moderation.
4. Data rules stack. PIPL (personal information), the Data Security Law, and cross-border data transfer requirements apply alongside the AI rules; training-data provenance and export of data out of China are frequent friction points.
5. Labelling is now mandatory and audited at the platform layer. Build explicit and implicit labelling into any China-facing generation pipeline.
6. Most foreign teams operate via a local entity or partner. Filing, hosting and moderation obligations in practice push toward a Chinese subsidiary or joint arrangement.

The through-line is that China's regime is operational and enforced today, built around registry filing, content control and labelling, rather than the risk-classification logic of the EU. For a foreign team, the first question is rarely 'which risk tier' and almost always 'is this offered to the Chinese public, and if so, is it filed and moderated to Chinese standards'.