Your Chinese partner says the product needs a CAC filing. Your lawyer mentions an MIIT licence, your investor asks about SAMR standards, and a hospital customer insists on NMPA registration. None of them is wrong, and none is describing the whole picture. China does not authorise AI through a single gate: authority is divided by function — content, connectivity, product safety, sector — and by level of government. For a foreign company the useful question is not “who regulates AI in China” in the abstract, but which body can stop your specific product, at which stage, and what it will ask for.
Who this guide is for
- Foreign AI companies testing whether a China launch is feasible and what approvals it implies.
- Investors and corporate development teams running regulatory diligence on a Chinese AI target.
- Legal, product and compliance leads already operating in China who need to assign an owner to each approval.
Stage: Explore. Read this before you commit to a structure or a launch date, because the regulator map usually changes the plan.
Why there is no single AI regulator
China regulates AI through the regulators it already had. There is no standalone AI ministry and, as of late 2026, no comprehensive AI statute in force. Two layers interact instead.
The horizontal layer applies across sectors: the Cybersecurity Law, the Data Security Law, the Personal Information Protection Law (PIPL), the Interim Measures for the Management of Generative AI Services, the Algorithm Recommendation Provisions, the Deep Synthesis Provisions and the AI-Generated Content Labelling Measures. These are administered mainly by the Cyberspace Administration of China (CAC), working with public security, industry and market regulators.
The vertical layer applies because of what the product is or where it is sold: medical software goes to the National Medical Products Administration (NMPA), vehicles to transport and industry authorities, financial models to financial regulators, listed companies to the China Securities Regulatory Commission (CSRC).
A third, softer layer is standardisation. Many AI standards are recommended (GB/T) rather than mandatory (GB), but Chinese customers, state-owned enterprises and local governments increasingly write them into tenders. Standards rarely give you permission; they decide whether you win the contract.
The result is that one product can face four agencies, and the agency that matters most depends on the feature carrying the most risk.
CAC: content, algorithms and data flows
The CAC, also known as the Cyberspace Administration, is the most visible AI regulator and usually the first one a foreign company meets. It operates through provincial and municipal offices that receive filings and act as the day-to-day interface.
Algorithm filing. Providers of algorithm recommendation services with public-opinion attributes or social mobilisation capacity must file with the local CAC office, typically within ten working days of starting to serve the public. The filing describes the algorithm, its purpose, data sources and self-assessment. It is a disclosure regime, not an approval of quality.
Generative AI filing and security assessment. Under the Generative AI Interim Measures, services provided to the public in China with public-opinion attributes or social mobilisation capacity must complete a security assessment and file with CAC. Materials go first to the provincial office, and the national administration publishes the list of filed services. Using a third-party model that is already filed does not automatically cover your application: the application itself may need separate registration.
Labelling. The AI-Generated Content Labelling Measures, effective from September 2025, require explicit labels that users can see and implicit labels embedded in files or metadata. They apply to generated text, images, audio, video and virtual scenes, and to distribution platforms as well as generators.
Data export. Cross-border transfer of personal information and important data is supervised by CAC. Depending on volume and sensitivity, a company may need a security assessment, a standard contract filing or certification. The 2024 cross-border data flow provisions raised thresholds and created exemptions, but the analysis remains fact-specific.
Content and security duties. Beyond filings, CAC enforces content moderation, user complaint handling, protection of minors and, for large platforms, cybersecurity review. These duties continue after launch and are the most common source of penalties.
MIIT: networks, devices and industrial policy
The Ministry of Industry and Information Technology (MIIT) is the industry and telecom regulator. It does not approve AI models as such, but it controls the pipes and many of the products.
- Internet access and apps. A commercial website or app serving users in China needs ICP filing; commercial internet information services need an ICP licence, a type of value-added telecom licence. Mobile apps must complete app filing. Without these, hosting and app-store distribution can be suspended.
- Radio and connected devices. Radio transmitting equipment needs type approval, often called SRRC approval, and telecom terminal equipment needs network access licensing.
- Compute and industrial policy. MIIT sets direction for data centres, compute hubs, AI plus manufacturing programmes and the robotics industry, and channels subsidies and pilot programmes through local bureaus.
- Intelligent connected vehicles. Vehicle product access and pilot programmes for intelligent connected vehicles run through MIIT together with transport, market regulation and public security authorities.
For hardware companies, MIIT often decides whether a device can legally connect, not whether its AI is any good.
SAMR, NMPA, MOFCOM and CSRC: product, sector and money
SAMR. The State Administration for Market Regulation covers standards, product quality, advertising, unfair competition and consumer protection. AI-specific touchpoints include compulsory certification for products inside the CCC catalogue, enforcement against misleading performance claims under the Advertising Law, and unfair competition cases involving data scraping or false labelling. SAMR also oversees the standards system that turns policy into procurement requirements.
NMPA. AI software that qualifies as a medical device — diagnostic imaging tools or clinical decision support, for example — needs medical device registration, usually Class II or Class III, supported by clinical evaluation. This is a product registration track that runs in parallel with any CAC filing.
MOFCOM. The Ministry of Commerce administers foreign investment through the negative list and information reporting, export controls on dual-use items, and the unreliable entity list. If your project involves investing in a Chinese AI company or moving controlled technology across the border, MOFCOM is in the room.
CSRC. AI companies listing in Shanghai, Shenzhen or overseas engage CSRC, including through the overseas listing filing regime. Data handling, model compliance and related-party arrangements typically surface during listing review.
The map at a glance
| Regulator | Core mandate | What triggers it | Typical first contact |
|---|---|---|---|
| CAC | Content, algorithms, data flows | Public-facing generative or recommendation service; outbound data | Provincial CAC office |
| MIIT | Telecom, devices, industrial policy | Website, app, radio device, connected vehicle | Provincial communications administration |
| SAMR | Standards, certification, advertising | CCC-listed product; consumer claims; unfair competition | Local market regulation bureau |
| NMPA | Medical devices | AI software with a medical purpose | NMPA or provincial drug regulator |
| MOFCOM | Investment and export control | Foreign investment; controlled technology | Provincial commerce department |
| CSRC | Securities and listing | IPO or overseas listing | CSRC filing system |
Putting the map to work
Central versus local
Beijing writes the rules; provincial and municipal governments decide how quickly you meet them. Local CAC offices receive filings and can return an incomplete package. Local industry bureaus control subsidies, pilot programmes and, in practice, how strictly standards are applied in public procurement. Cities such as Beijing, Shanghai and Shenzhen compete for AI companies and publish their own support policies, which can help but also add reporting obligations. A common mistake is to negotiate only with the local government offering incentives, then discover that the provincial CAC office takes a different view of the same product.
A product’s path through the system
A consumer generative AI app typically meets: company establishment or a local partner; ICP filing or licence plus app filing through the MIIT system; algorithm filing and, if it serves the public, generative AI filing with a security assessment through CAC; labelling implementation; a data export analysis if any data leaves China; advertising review for performance claims through SAMR; and ongoing content duties. An AI medical device adds NMPA registration and hospital procurement requirements. An AI-enabled robot sold to factories adds product safety, radio approval if wireless, and certification demands written into the customer contract.
How to identify your lead regulator
Ask four questions. First, does the product put content in front of the Chinese public? If yes, CAC leads. Second, is it embedded in a regulated product such as a vehicle, a medical device or a radio device? Then the sector regulator leads and CAC follows. Third, who is the customer? Consumer sales attract SAMR and CAC; hospitals attract NMPA; state-owned enterprises import their own certification demands. Fourth, where does the data go? Any outbound transfer brings CAC back in. Where two regulators both claim jurisdiction, the practical answer is usually that both must be satisfied, and sequencing matters more than hierarchy.
What to do next
- Write a one-page product description stating what the model does, who sees the output, what data it touches and where that data is stored.
- Map each feature to a regulator and mark whether the requirement is a permission, a filing or a standard.
- Confirm whether you need a Chinese entity, and in which province it should sit, before drafting any filing.
- Check the filing status of any base model or API you intend to use, and whether your application needs separate registration.
- Build labelling and content moderation into the product specification rather than the launch checklist.
- Run a data export analysis early; retrofitting data localisation is expensive.
- Ask local counsel to pre-clear the filing package with the provincial CAC office before formal submission.
Rules in this area change frequently. This article reflects the position as of late 2026 and should be confirmed with counsel before you rely on it for a filing or launch decision.
Frequently Asked Questions↑ Back to article
Is there a single AI regulator in China?
No. China regulates AI through existing sector regulators plus cross-cutting rules on content, data and algorithms, with the Cyberspace Administration of China (CAC) as the most visible authority. Product-specific regulators such as NMPA for medical devices and MIIT for telecom and devices also apply, so a single AI product can face several agencies at once.
Do we need a Chinese entity before filing with CAC?
In practice, yes. Filings and security assessments are submitted by the entity that provides the service in China, normally a Chinese company, and are lodged with the provincial CAC office where that entity operates. A foreign company can license technology into China, but the local operating company usually carries the filing obligation.
What does MIIT regulate that CAC does not?
MIIT controls market access for networks and devices: ICP filing and licensing for websites and apps, app filing, radio type approval for wireless equipment and network access licensing for telecom terminals. It also sets industrial policy for compute, robotics and intelligent connected vehicles. CAC focuses on content, algorithms and data flows rather than connectivity.
Which regulator handles AI medical software?
The National Medical Products Administration (NMPA) treats AI software with a medical purpose as a medical device, usually requiring Class II or Class III registration supported by clinical evaluation. A CAC filing may still be needed for any public-facing or data-intensive element, so the two tracks run in parallel rather than replacing each other.
Who reviews cross-border data transfers from China?
CAC supervises outbound transfers of personal information and important data, working through provincial offices. Depending on volume, sensitivity and the nature of the data, a company may need a security assessment, a standard contract filing or certification. The analysis is fact-specific and should be documented before data starts to flow.
Does SAMR approve AI products before they can be sold?
Only for products inside the compulsory certification catalogue, where CCC certification is required before sale. Otherwise SAMR's role is standards, product quality, advertising and unfair competition enforcement, which shapes how you may describe and market an AI product rather than whether you may sell it.
How do we identify our lead regulator?
Ask whether the product puts content in front of the Chinese public, whether it is embedded in a regulated product such as a vehicle or medical device, who the customer is, and where the data goes. The regulator attached to the highest-risk feature usually leads, but where two agencies claim jurisdiction both normally have to be satisfied.