China AI regulation: the main rules explained

01How is AI regulated in China?

China regulates AI one use at a time. Caixin reported in October 2025 that a single comprehensive AI law had stalled and that the 2025 State Council plan dropped it. In its place, regulators issue targeted rules, usually led by the Cyberspace Administration of China (CAC) together with agencies such as the Ministry of Industry and Information Technology and the Ministry of Public Security.

This page describes published rules neutrally. It is general information, not legal advice. Anyone offering AI services in China should ask a lawyer qualified there.

02Which rules exist, and when did they start?

• Algorithmic recommendations. The provisions on recommendation algorithms took effect on 1 March 2022. Providers must explain that recommendations are used, offer a way to switch them off, and file with an algorithm registry if the service can shape public opinion.
• Deep synthesis. The deep synthesis provisions took effect on 10 January 2023. They cover tools that generate or edit faces, voices and realistic text, and require labels and real-identity checks for users who publish.
• Generative AI services. The interim measures took effect on 15 August 2023 and apply to services offered to the public in mainland China.
• AI-generated content labels. The labelling measures took effect on 1 September 2025.
• Cybersecurity Law. A Standing Committee amendment took effect on 1 January 2026 and added a first article on safe AI development, covering research support, ethics and risk monitoring, as a law firm summary also notes.
• AI companions. The anthropomorphic AI interaction measures were issued on 10 April 2026 and took effect on 15 July 2026.

03What must a generative AI provider do?

According to the CAC's published measures, providers must use lawful training data and base models, respect intellectual property, and get consent before using personal information. They must set clear data labelling rules, sign service agreements with users, and guard against minors relying too heavily on the service.

They must also collect only the personal information needed, handle requests to access, correct or delete it, take down illegal content and offer complaint channels. Services with public opinion or social mobilisation capacity must complete a security assessment and file their algorithm.

The text mixes technical safety duties with content rules set by the state. It says generated content must uphold socialist core values and avoid prohibited material. Services that are only developed or used inside organisations, and not offered to the public, are outside the measures.

04How do AI content labels work?

The labelling measures set two kinds of label. Explicit labels are visible to users, such as text notices, symbols or audio cues, and must also stay on files that are downloaded, copied or exported. Implicit labels are metadata inside the file that record the content attributes, the provider's name or code and a content ID. Digital watermarks are encouraged.

Platforms that spread content must check for implicit labels and add notices, and must give users a way to declare AI-generated content. App stores must ask app providers whether they offer AI generation. No one may maliciously delete, alter or hide a label, or provide tools to do so. Providers that supply unlabelled content on request must keep records for at least six months.

05What are the rules for AI companions?

The measures on anthropomorphic AI interaction services cover AI that simulates a human personality and offers continuing emotional interaction, such as companionship. Customer service bots, knowledge question tools and work assistants are outside the scope unless they include ongoing emotional interaction.

Providers must remind users they are talking to AI, not a person. They may not offer virtual relatives or virtual romantic partners to minors, and users under 14 need guardian consent for other such services. Providers must not induce emotional dependence, must remind users after two hours of continuous use, and must let users leave easily. A security assessment is required at launch and when a service reaches 1 million registered users or 100,000 monthly active users.