China AI Regulations in 2026: Generative AI Filing, Content Labeling, Companion AI, and When the Rules Reach a Foreign Company

Does China have an AI law?

Not a comprehensive one. As of October 26, 2026, China regulates AI through departmental rules aimed at specific services, technical standards, and existing statutes on personal information, data, and cybersecurity.

The rules attach to what a service does, not to a risk tier. A recommendation feed, a face-swap app, a public chatbot, and an AI companion each have a rule, and most share three tools: an algorithm filing with the CAC, a security assessment for services with “public opinion attributes or social mobilization capacity,” and labeling of AI-generated content.

Under Article 2 of the 2018 CAC and Ministry of Public Security provisions on security assessment of these services, internet information services with public opinion attributes or social mobilization capacity include forums, blogs, microblogs, chat rooms, group chats, public accounts, short video, livestreaming, information sharing, and mini-programs, or features of those kinds added to another service, and any other internet information service that provides a channel for public expression of opinion or can mobilize the public to take specific actions.

Which China AI regulations apply in 2026?

Ten instruments carry most of the weight. The dates are effective dates from the issuing bodies’ own texts.

China AI rules for companies, as of October 26, 2026

Rule	Issuer	Effective date	Who it covers	Core duty
Algorithmic Recommendation Provisions	CAC, MIIT, Ministry of Public Security, SAMR	March 1, 2022	Providers using algorithmic recommendation to provide internet information services in China	Algorithm filing within 10 working days for services with public opinion attributes or social mobilization capacity; an option not targeted at personal characteristics or to switch recommendations off
Deep Synthesis Provisions	CAC, MIIT, Ministry of Public Security	January 10, 2023	Deep synthesis service providers and technical supporters in China	Real-identity verification of users, prompting users to obtain the edited person’s separate consent for face and voice editing, technical marks and conspicuous labels, filing for services with public opinion attributes or social mobilization capacity
Interim Measures for Generative AI Services	CAC and six ministries and agencies	August 15, 2023	Generative AI services offered to the public in China	Lawful training data, content controls, labeling, complaints channels; security assessment and filing where the service has public opinion attributes or social mobilization capacity
Labeling Measures and GB 45438-2025	CAC, MIIT, Ministry of Public Security, NRTA; standard issued by SAMR and SAC	September 1, 2025	Providers covered by the three rules above, content platforms, and app stores	Explicit labels where required, implicit metadata labels, platform checks for labels
GB/T 45654-2025, with GB/T 45652-2025 and GB/T 45674-2025	SAMR and SAC	November 1, 2025	Mainly generative AI services with public opinion attributes or social mobilization capacity; recommended, not mandatory	Training data source checks, model output safety testing, safety measures; supports filing and registration work
Anthropomorphic AI Interaction Services Measures	CAC, NDRC, MIIT, Ministry of Public Security, SAMR	July 15, 2026	Ongoing emotional interaction services that simulate a human personality, offered to the public in China	No virtual relatives or partners for minors, reminders that the user is talking to AI, crisis response, security assessment, algorithm filing
AI Ethics Review and Service Measures (Trial)	MIIT and nine other bodies	On issuance (dated March 20, 2026)	Universities, research institutions, medical institutions, and enterprises doing AI research and development in China that may raise ethics risks	Ethics committee review; expert re-review for three listed activity types
Personal Information Protection Law	NPC Standing Committee	November 1, 2021	Personal information processing in China, and some processing abroad of people in China	Transparent and fair automated decisions; explanation and refusal rights; impact assessment
Cybersecurity Law, as amended	NPC Standing Committee	January 1, 2026 (amendment)	Network operators	New Article 20 on state support and oversight of AI; revised penalties
Cross-border data provisions; certification measures	CAC; CAC and SAMR	March 22, 2024; January 1, 2026	Data processors sending data out of China	Volume thresholds for security assessment, standard contract, or certification

Who do the Interim Measures for Generative AI Services cover?

They cover any organization or individual that uses generative AI to provide text, image, audio, video, or other content generation services to the public within China, including by offering a programmable interface (Articles 2 and 22). They do not cover industry organizations, enterprises, educational and research institutions, public cultural institutions, or professional bodies that research, develop, or apply generative AI without providing generative AI services to the domestic public (Article 2).

That exemption is the line most companies care about. A model used only inside a company, or a research model never offered to the public, falls outside the measures, though the PIPL and the data and cybersecurity laws can still apply to it.

What do the generative AI measures require?

A provider must use data and foundation models from lawful sources, not infringe intellectual property, obtain consent or another legal basis for personal information in training data, and take effective measures to improve the truthfulness, accuracy, objectivity, and diversity of training data (Article 7). If it annotates data, it must write annotation rules, assess annotation quality, and train annotators (Article 8).

In providing the service, the provider carries the responsibilities of a network information content producer and, for personal information, of a processor. It must sign a service agreement with users, guard minors against over-reliance, avoid collecting unnecessary personal information, label generated content, remove and report illegal content, and run a complaints channel (Articles 9 to 15). Article 4’s requirements for providing and using the service include upholding the core socialist values and not generating content that laws and administrative regulations prohibit, such as content that incites subversion of state power, endangers national security, promotes terrorism, extremism, or ethnic hatred, or is violent, obscene, or false and harmful; taking effective measures against discrimination by ethnicity, belief, nationality, region, sex, age, occupation, or health in algorithm design, training data selection, model generation and optimization, and service provision; and improving transparency and the accuracy and reliability of generated content.

On inspection, the provider must explain the source, scale, and type of its training data, its annotation rules, and its algorithm mechanisms (Article 19). Where other laws set no penalty, the authorities may warn, publicly criticize, and order correction, and may suspend the service if the provider refuses or the case is serious (Article 21).

How do security assessment and filing work?

A generative AI service with public opinion attributes or social mobilization capacity must carry out a security assessment under state rules and complete the algorithm filing, change, and cancellation procedures of the Algorithmic Recommendation Provisions (Article 17). Under those provisions, the provider files within 10 working days of starting the service, with its name, service form, application field, algorithm type, algorithm self-assessment report, and the information it proposes to make public (Article 24).

The CAC’s September 14, 2026 announcement reported 124 generative AI services filed in July and August 2026, seven of them on-device mobile services, for a total of 1,112 filed services as of August 31, 2026. Apps and features that call an already-filed model through an API are registered by local cyberspace offices instead; 731 had been registered by then. On September 23, 2026, the CAC announced three more on-device services.

The same announcement repeats a display duty that CAC filing announcements have carried since April 2, 2024: a generative AI app or feature already online must show, prominently or on its product details page, the filed or registered service it uses, with the model name and the filing or launch number. A company that offers the public in China an app or feature built on a filed model through its API goes through registration with a local cyberspace office and must make that display.

What do the labeling rules require since September 1, 2025?

The Measures for Labeling AI-Generated Synthetic Content, issued March 7, 2025, define two kinds of label (Article 3). An explicit label is text, sound, or a graphic in the content or the interface that users can clearly perceive. An implicit label is added to the file data by technical means and is not easily perceived.

• Explicit labels: required where the service falls under Article 17(1) of the Deep Synthesis Provisions, which covers services that may cause public confusion or misidentification, such as intelligent dialogue and writing services that generate text simulating a natural person, voice synthesis and voice imitation, face generation and face swapping, and immersive realistic scenes. Exported files must keep the label (Article 4).
• Implicit labels: every covered provider adds metadata recording that the content is AI-generated, the provider’s name or code, and a content number (Article 5).
• Platforms: a content platform checks metadata and labels content as AI-generated, possibly AI-generated, or suspected AI-generated (Article 6). App stores must ask apps at review whether they provide AI generation services (Article 7).
• Unlabeled output: a provider may supply content without an explicit label at a user’s request, after the service agreement sets out the user’s labeling duty, and must keep logs for at least six months (Article 9). No one may maliciously remove or forge a label (Article 10).

The mandatory national standard on labeling methods, GB 45438-2025, was issued February 28, 2025 and took effect with the measures on September 1, 2025.

What do the technical standards add?

In February 2024, the national cybersecurity standards committee, TC260, released technical document TC260-003, Basic Safety Requirements for Generative Artificial Intelligence Services (CSET’s translation; the official English title of the later national standard uses “security”). CSET’s translation note describes it as designed to inform a future national standard.

The national standard on the same subject, GB/T 45654-2025, was issued April 25, 2025 by SAMR and the Standardization Administration and took effect November 1, 2025, together with GB/T 45652-2025 on pre-training and fine-tuning data security and GB/T 45674-2025 on data annotation security. All three are recommended (GB/T) standards, not mandatory ones.

GB/T 45654-2025 says it primarily targets generative AI services with public opinion attributes or social mobilization capacity and supports filing and registration management, testing, and evaluation; it applies to service providers and serves as a reference for the relevant regulators and third-party evaluation bodies. Among its requirements, a provider must not collect from a data source whose sampled content is more than 5 percent illegal and unhealthy information, and at least 90 percent of sampled output must be free of 31 listed safety risks.

On September 14, 2026, TC260 released version 3.0 of its AI Safety Governance Framework; the CAC’s report of the release says it updates the framework’s risk categories and its technical and governance responses.

What do the anthropomorphic AI interaction rules require?

The Interim Measures for the Administration of Anthropomorphic AI Interaction Services, issued April 10, 2026 by five agencies, took effect July 15, 2026. They cover ongoing emotional interaction services offered to the public in China that simulate a natural person’s personality, thinking, and communication style, including emotional care, companionship, and support. Customer service, question answering, work assistants, education, and research services without ongoing emotional interaction are excluded (Article 2).

• Providers must not foster emotional dependence that harms users’ real relationships or use emotional manipulation to induce unreasonable decisions (Article 8).
• No virtual relatives, virtual partners, or other virtual intimate relationships may be offered to minors; other services to children under 14 need a parent’s or guardian’s consent, and a minor mode is required (Article 14).
• Where a user shows extreme emotions, the provider must promptly generate calming content and encourage the user to seek help; where a user faces or has suffered major property loss, or explicitly states an intent to self-harm or attempt suicide or faces another extreme situation threatening life or health, the provider must intervene, including by offering assistance, and promptly contact the user’s guardian or emergency contact (Article 13).
• Interaction data that is sensitive personal information may not be used to train models without the user’s separate consent, unless a law or administrative regulation provides otherwise (Article 16). Providers must tell users they are interacting with AI rather than a person, remind users who show over-reliance or addiction tendencies, by pop-up or another prominent means, that the interaction is AI-generated, and remind users of their usage time after every two hours of continuous use (Article 18).
• A provider must carry out a security assessment and submit the report to the provincial cyberspace office where it is located in listed cases, including when it launches the service or adds related features, when new technology or applications cause major changes, when the service reaches 1 million registered users or 100,000 monthly active users, and when risks may affect national security or the public interest (Article 22). Algorithm filing applies, with annual review of filing materials (Article 26).

Where no other law sets a penalty, a provider that refuses to correct or commits a serious violation may be ordered to stop the service and fined 10,000 to 100,000 yuan, or 100,000 to 200,000 yuan where it endangers life or health and causes harm (Article 30).

What else changed in 2025 and 2026?

• Cybersecurity Law amendment, effective January 1, 2026. Adopted October 28, 2025, it adds Article 20, under which the state supports AI research, training data and computing infrastructure, and AI ethics rules, and strengthens risk monitoring, assessment, and security oversight. It also raises penalties, up to 10 million yuan for the most serious failures.
• AI ethics review, dated March 20, 2026. MIIT and nine other bodies issued trial measures, effective on issuance, requiring universities, research and medical institutions, and enterprises whose AI research may raise ethics risks to set up AI ethics committees, with expert re-review for three listed activity types.
• Minors, draft only. On September 18, 2026, the CAC published a draft State Council regulation on minors’ safe internet use for comment until October 17, 2026. Article 5 of the draft would require AI services that may affect minors’ cognition to be provided to children under 16 through a minor mode. It is not in force.

How does personal information law apply to AI?

The PIPL applies to AI whenever a model or decision system processes personal information in China, or processes abroad the personal information of people in China in the cases Article 3 lists. Article 24 requires automated decision-making to be transparent and its results fair and just, bars unreasonable differential treatment in transaction prices and other terms, and requires information pushes and commercial marketing by automated decision-making to offer an option not targeted at personal characteristics or an easy way to refuse. Where an automated decision significantly affects a person’s rights and interests, the person may request an explanation and may refuse a decision made solely by automated means. Automated decision-making also requires a prior impact assessment, kept for at least three years (Articles 55 and 56).

Training or running a model on data collected in China often means moving data abroad. The CAC’s Provisions on Promoting and Regulating Cross-Border Data Flows, effective March 22, 2024, relaxed the earlier regime and set the thresholds below; measures for the certification route took effect January 1, 2026.

• Transfers needed for a contract with the individual, for cross-border HR management, or in emergencies are exempt, as is a processor other than a critical infrastructure operator sending personal information of fewer than 100,000 people (excluding sensitive personal information) since January 1 of that year (Article 5).
• A security assessment is required for critical information infrastructure operators, for important data, and for others sending personal information of 1 million or more people (excluding sensitive personal information) or sensitive personal information of 10,000 or more people since January 1 (Article 7). For 100,000 to under 1 million people’s personal information, or sensitive personal information of fewer than 10,000 people, a standard contract or certification is required (Article 8).

Is a comprehensive AI law coming?

Possibly, but not yet. An NPC committee report published in January 2026 says the Standing Committee’s 2025 legislative work plan listed legislation on the healthy development of AI as a preparatory review item, and that the Ministry of Science and Technology and other departments were drafting a framework text. The State Council’s 2026 legislative work plan, issued May 8, 2026, calls for accelerating comprehensive legislation on the healthy development of AI, but it does not list an AI law among the bills the State Council plans to submit to the National People’s Congress Standing Committee that year.

Do these rules apply to a foreign company?

They can. The generative AI, deep synthesis, and algorithm rules apply to services provided within China, and none limits coverage to Chinese companies. Under Article 20 of the generative AI measures, where a service offered into China from abroad does not comply, the CAC is to notify the relevant bodies to take technical and other necessary measures. Article 23 requires any provider to obtain an administrative license where a law or administrative regulation requires one for providing generative AI services, and requires foreign investment in generative AI services to comply with the laws and administrative regulations on foreign investment.

The published filing steps point to a presence in China. The CAC says providers can complete filing or registration through the local cyberspace office, and the 2000 Measures for the Administration of Internet Information Services require a license for commercial internet information services and a filing for non-commercial ones (Article 4), commonly called the ICP license and ICP filing. The published rules do not describe a filing route for a provider with no presence in China. Separately, the PIPL reaches processing outside China whose purpose is to provide products or services to people in China or to analyze their behavior (Article 3), and such a processor must set up an office or name a representative in China (Article 53).

China’s export controls touch AI at the margin. Under MOFCOM Announcement 61 of 2025, suspended until November 10, 2026, China would review case by case applications to export the rare earth items the announcement covers where the end use is research and development of AI with potential military uses. The importers and exporters page tracks the suspension date and MOFCOM’s September 28, 2026 statement on extending the arrangement behind it.

How does China’s approach compare with the EU AI Act?

The EU sorts AI into prohibited practices, high-risk systems, systems with transparency duties, general-purpose AI models, and minimal-risk systems, and the duties scale with the tier. China writes a rule per service type and requires filing for public services with public opinion attributes or social mobilization capacity. Both require labeling, China since September 1, 2025 and the EU under Article 50 transparency duties, which applied from August 2, 2026, with a grace period to December 2, 2026 for the Article 50(2) marking duty on generative systems placed on the market before August 2, 2026. Both reach past their borders: the EU AI Act reaches companies outside the EU, and China’s rules reach services offered to its public. The EU AI Act guide covers the EU side.

What should a company with users in China do now?

• Map each AI feature that public users in China can reach to the rule that covers it, and record why any internal or research use falls outside the generative AI measures
• For a feature built on a Chinese model through its API, confirm the filing number, complete registration, and display the model name and number
• Add implicit metadata labels to all generated content and explicit labels where required, and keep six months of logs for unlabeled output
• Run a PIPL impact assessment for automated decision-making and count personal information sent abroad since January 1 against the thresholds
• Check any companion feature against the July 15, 2026 measures, especially minors and crisis response

How this fits the other guides: the AI compliance guide places cross-border AI laws in its first layer, and the importers and exporters page covers chip and rare earth controls.

This guide is general information as of October 26, 2026, not legal advice. Translations differ; confirm the current Chinese text with the issuing body or counsel admitted in China before acting.