China's AI rules

What it is

China does not regulate AI through one statute. It regulates specific uses through administrative regulations issued mostly by the Cyberspace Administration of China (CAC), each targeting a type of service as it became commercially significant. Read together, they form a layered regime.

• Algorithmic Recommendation Provisions. The Internet Information Service Algorithmic Recommendation Management Provisions, issued by the CAC, MIIT, the Ministry of Public Security and SAMR, took effect on 1 March 2022. They cover recommendation, ranking, retrieval, scheduling and generative algorithms used to provide internet information services.
• Deep Synthesis Provisions. The Internet Information Service Deep Synthesis Management Provisions took effect on 10 January 2023. They cover technologies that generate or edit text, images, audio, video and virtual scenes.
• Generative AI Interim Measures. The Interim Measures for the Management of Generative Artificial Intelligence Services, issued by the CAC with six other agencies, took effect on 15 August 2023.
• Labelling Measures. The Measures for Labelling AI-Generated Synthetic Content, from the CAC, MIIT, the Ministry of Public Security and the National Radio and Television Administration, took effect on 1 September 2025, together with the mandatory national standard GB 45438-2025 on labelling methods.
• Anthropomorphic interaction services. Interim Measures for AI anthropomorphic interaction services, issued by five departments on 10 April 2026, took effect on 15 July 2026.
• Cybersecurity Law, Article 20. The NPC Standing Committee amended the Cybersecurity Law on 28 October 2025, effective 1 January 2026. The new Article 20 commits the state to support AI research and infrastructure, improve AI ethics norms and strengthen risk monitoring, assessment and safety supervision.

Alongside the binding rules sits the AI Safety Governance Framework, a voluntary technical document from TC260 prepared under CAC guidance. Version 2.0 (September 2025) sorts risks into inherent technical risks, application risks and derivative risks, and sets out governance measures across the lifecycle. Version 3.0 followed on 14 September 2026, keeping the same structure of risk classification, technical countermeasures and comprehensive governance while updating the risk categories and measures.

Who it applies to

The binding rules apply to services provided within mainland China. The Generative AI Interim Measures cover services that use generative AI to provide content to the public in mainland China. Research and development by companies, universities and research bodies that is not offered to the domestic public is outside their scope. The Labelling Measures apply to generative and synthetic content service providers and to platforms that distribute content.

For organisations outside China, these rules come into play in two ways. You may operate a service that reaches the public in mainland China. More often, you buy AI from, or build on, a Chinese model or platform provider that is itself subject to these rules. That makes this page mainly a reference for vendor due diligence.

What it requires

Across the instruments, the recurring duties are these.

• Algorithm filing and security assessment. Providers of services with public opinion attributes or social mobilisation capacity must file their algorithms with the CAC and complete a security assessment. The duty originates in the Algorithmic Recommendation Provisions and is applied to deep synthesis and generative AI services.
• Training data. Generative AI providers must use data and foundation models from lawful sources, respect intellectual property, have consent or another lawful basis for personal information and take steps to improve data quality. Data annotation needs clear rules, quality checks and trained annotators.
• Content responsibility. Providers are responsible for the content their service generates. They must stop, correct and report illegal content and protect users’ inputs and usage records.
• Labelling. Explicit labels must be visible on generated text, audio, images, video and virtual scenes, including in downloads and exports. Implicit labels must be embedded in metadata. Distribution platforms must check for labels and add notices. Removing or forging labels is prohibited.
• User rights. Recommendation services must let users switch off personalised recommendation or delete their profile tags. Providers must offer accessible complaint and reporting channels and respond to them.
• Identity and biometrics. Deep synthesis services must verify users’ real identity, and editing a person’s face or voice requires that person’s separate consent.

Key dates

China’s rules arrived instrument by instrument. Algorithmic recommendation on 1 March 2022, deep synthesis on 10 January 2023, generative AI on 15 August 2023, labelling on 1 September 2025, the Cybersecurity Law’s AI article on 1 January 2026 and anthropomorphic interaction services on 15 July 2026. TC260 has published a new version of the AI Safety Governance Framework each September since 2024.

How it relates to other frameworks

The labelling duties sit close to the transparency obligations in the EU AI Act and the generative AI notice duties in South Korea’s AI Basic Act, though China’s rules add metadata requirements and platform checks that most other regimes leave to standards. Filing and security assessment have no direct equivalent in Western frameworks. The AI Safety Governance Framework’s lifecycle measures overlap with the risk management practices in ISO/IEC 42001 and the NIST AI RMF, which gives organisations a common vocabulary when they question a Chinese supplier.

For due diligence on a Chinese AI supplier, useful questions are whether its service has completed algorithm filing and any required security assessment, how it labels generated content and embeds metadata, where its training data came from and how it handles your inputs.

A note on this page

This is Aicura’s reading of China’s AI rules, written as a reference. It is not legal advice and it is not the rules themselves. The authoritative texts are in Chinese. Read the primary sources from the CAC and TC260, and take advice on how they bear on your organisation.