🇨🇳China

Detailed overview

China does not regulate artificial intelligence through one single AI Act. Instead, it has several binding AI-specific rules that apply to different AI use cases, especially public-facing generative AI, algorithmic recommendation systems, deep synthesis technologies and AI-generated content labelling. These rules are mainly issued by the Cyberspace Administration of China and other national authorities.

Generative AI services

China's Interim Measures for the Management of Generative Artificial Intelligence Services apply to the use of generative AI technologies to provide services to the public in China that generate text, images, audio, video or other content. Internal research, development and non-public use are treated differently. A generative AI service means an AI service that can generate new content based on models, algorithms and data. The rules entered into force on 15 August 2023.

Providers of public-facing generative AI services must comply with content, safety, data and user-protection duties. They must not generate illegal content, must respect socialist core values, must prevent discrimination, must protect intellectual property and business secrets, and must respect personal information and personal rights. Providers must also take measures to improve the truthfulness, accuracy, objectivity and reliability of generated content.

Training data must be lawful. Providers must use lawful data sources and foundation models, must not infringe intellectual property rights, must obtain consent or another lawful basis where personal information is involved, and must take measures to improve training-data quality. Where a generative AI service has public opinion attributes or social mobilisation capacity, the provider must complete a security assessment and algorithm filing under China's algorithmic recommendation rules.

Generative AI providers must also manage user relationships. They must clarify service rules, protect users' input information and usage records, avoid unnecessary collection of personal information, respond to complaints and correction requests, and take measures against illegal content. Where illegal content is discovered, the provider must stop generation or transmission, remove or rectify the content, optimise the model where necessary and report to the competent authority where required.

Algorithmic recommendation

China also regulates algorithmic recommendation services. These rules apply to recommendation technologies such as personalised push, ranking and selection, search filtering, dispatch decision-making, generative or synthetic algorithms and other algorithmic services used to provide information to internet users. Algorithmic recommendation providers must maintain algorithm-security systems, conduct algorithm reviews, protect data and personal information, and avoid using algorithms in ways that endanger national security, public order or the lawful rights of individuals and organisations.

Users must be informed when algorithmic recommendation services are used. Providers must disclose the basic principles, purposes and main operating mechanisms of the algorithmic service in an appropriate way. Users must be offered a way to turn off personalised recommendation or receive non-personalised options, and they must be able to select or delete user tags used for recommendations. Where an algorithmic service has a major impact on a user's rights or interests, the provider must explain the decision and provide a complaint channel.

Deep synthesis

China's Provisions on the Administration of Deep Synthesis Internet Information Services regulate technologies that use deep learning, virtual reality and similar synthesis algorithms to generate or significantly alter text, images, audio, video or virtual scenes. These rules entered into force on 10 January 2023. Deep synthesis providers must verify users, review input and output content, protect data and personal information, prevent fraud, maintain complaint channels and take emergency measures where unlawful or harmful information is found.

Deep synthesis services must also use labels. Providers must add invisible technical identifiers and, where content may confuse or mislead the public, visible or prominent labels. Separate consent is required where face, voice or other biometric information is edited or used in certain deep synthesis functions. Some face, voice and special-scene generation tools also require a security assessment.

Content labelling

China has adopted separate rules on AI-generated and synthetic content labelling. These rules require explicit and implicit labels for AI-generated or AI-synthetic content such as text, images, audio, video and virtual scenes. Explicit labels are visible or perceptible to users, while implicit labels may be embedded in metadata or technical identifiers. The rules also prohibit malicious deletion, tampering, forging or hiding of AI labels.

Penalties

Penalties depend on the breached law. The generative AI rules refer to enforcement under China's Cybersecurity Law, Data Security Law, Personal Information Protection Law and other applicable laws. Where no separate law applies, authorities may issue warnings, order corrections, criticise the provider, require suspension of services or take other administrative measures. Serious breaches may also trigger administrative or criminal liability.

Practical requirements & details

Sourced from official CAC instruments: Interim Measures for the Management of Generative AI Services (effective 15 Aug 2023), the Provisions on the Administration of Deep Synthesis Internet Information Services (effective 10 Jan 2023), the Provisions on the Administration of Algorithmic Recommendations (effective 1 Mar 2022), and the Measures for the Labelling of AI-Generated and Synthetic Content (joint CAC/MIIT/MPS/NRTA, effective 1 Sept 2025).

Generative AI Interim Measures (公開 services in China)

• Apply to public-facing generative AI services within China (internal R&D and non-public use is treated more lightly).
• Content rules: must respect socialist core values; no content that incites subversion, undermines national unity, promotes terrorism/extremism, ethnic discrimination, violence, obscenity, or false/harmful information.
• Training data: lawful sources; no IP infringement; consent or other lawful basis for personal information; measures to improve data quality, truthfulness, accuracy, objectivity and diversity.
• Where the service has public-opinion attributes or social-mobilisation capacity: complete a security assessment and an algorithm filing with CAC before launch.
• Identify and label generated content per the labelling rules; respond to user complaints; rectify/stop transmission of illegal content and report to authorities.

Deep Synthesis Provisions

• Apply to services using deep learning, virtual reality and other synthesis algorithms to generate or significantly alter text, images, audio, video or virtual scenes.
• Mandatory real-name verification of users.
• Content review of inputs and outputs; protect personal information; data-security measures.
• Add technical (implicit) identifiers to outputs; add visible (explicit) labels where content could mislead or confuse the public.
• Editing or use of face, voice or other biometric information: separate informed consent of the data subject required.
• Some face/voice/special-scene generation tools require a separate security assessment.

Algorithmic Recommendation Provisions

• Cover personalised push, ranking and selection, search filtering, dispatch decision-making, generative or synthetic algorithms and other algorithmic services.
• Establish algorithm-security management systems, conduct algorithm audits and protect data and personal information.
• Inform users of basic principles, purpose and main operating mechanisms.
• Provide an opt-out from personalised recommendations (or non-personalised alternative) and allow users to select or delete tags used for recommendations.
• Where the service has major impact on user rights or interests: explanation and complaint channel required.
• Algorithmic services with public-opinion or social-mobilisation attributes: complete algorithm filing with CAC.

AI Content Labelling Measures (Sept 2025)

• Apply to AI-generated/synthetic text, images, audio, video and virtual scenes.
• Explicit labels (visible/audible) on the content itself and around output.
• Implicit labels embedded in metadata or as technical identifiers (provenance markers).
• Prohibits malicious deletion, tampering, forging or hiding of AI labels.
• Platforms must verify, flag, propagate and preserve labels across distribution chains.

Cross-cutting laws and penalties

• Generative AI rules cross-reference the Cybersecurity Law, Data Security Law, and Personal Information Protection Law (PIPL) for enforcement.
• CAC may issue warnings, order corrections, publicly criticise the provider, suspend services, and seek further administrative or criminal penalties.
• PIPL fines for serious personal-information breaches: up to RMB 50,000,000 or 5% of preceding-year turnover.