China’s data and AI rules: what global companies need to know.

The three laws and the regulator

The Cybersecurity Law (2017) covers network security and the operators of “critical information infrastructure”, meaning systems whose failure could harm national security or the public interest. Amendments in force since 1 January 2026 add express references to AI and raise maximum fines, up to RMB 10 million in the most serious cases.

The Data Security Law (September 2021) requires organisations to classify data and gives special protection to “important data”, a category linked to national security and the economy. The Personal Information Protection Law, or PIPL (November 2021), is China’s equivalent of the GDPR.

The Cyberspace Administration of China (CAC) leads on all three, working with sector regulators. It also issues the more detailed rules on transfers, algorithms, generative AI and labelling discussed below.

Does PIPL apply to a company outside China?

PIPL applies to processing inside China. It also applies to processing abroad where the purpose is to provide products or services to people in China, or to analyse or assess their behaviour. This is similar to the GDPR’s reach, but enforcement can feel more direct.

A foreign e-commerce site shipping to customers in China, or an app tracking the behaviour of users in China, can be covered without any local entity. An offshore company in scope must set up a dedicated office or appoint a representative in China and report that person’s details to the regulator.

Penalties for serious violations reach RMB 50 million or 5% of the previous year’s turnover. Regulators can also suspend services, and responsible managers can face personal fines.

How PIPL differs from the GDPR

PIPL lists the lawful bases for processing, but there is no “legitimate interests” basis. In practice, consent carries more weight than in Europe, though contract performance, HR management and legal obligations are also recognised.

PIPL also uses “separate consent”, meaning specific consent for a specific activity rather than a box ticked within general terms. It is required for processing sensitive data, sharing data with another controller, publishing personal information, transferring it abroad, and using images from cameras in public places for anything other than public security.

• Sensitive data includes biometrics, health, financial accounts, location tracking and all data about children under 14
• Impact assessments are required before sensitive processing, automated decision-making, sharing and transfers abroad
• Compliance audits have been mandatory since 1 May 2025, at least every two years for those handling data on more than 10 million people

Cross-border transfers after the 2024 reforms

Sending personal information out of China needs one of three routes: a CAC security assessment, the standard contract (a government-issued template filed with the regulator) or certification by an accredited body. Rules on certification took effect on 1 January 2026, completing the set.

Rules issued in March 2024 raised the thresholds and created exemptions. Which route applies depends mainly on how many people’s data you export in a calendar year, counted from 1 January.

• Under 100,000 people (non-sensitive data): no formal mechanism, though consent and assessment duties still apply
• 100,000 to 1 million people, or fewer than 10,000 people’s sensitive data: standard contract or certification
• Over 1 million people, over 10,000 people’s sensitive data, important data, or any export by an infrastructure operator: security assessment
• Exempt: transfers needed to perform a contract with the individual, such as cross-border shopping or travel booking, and transfers needed for HR management under lawful employment policies

Important data and data localisation

“Important data” is separate from personal information. It can include industrial, mapping, health or other data that could affect national security if leaked. Under the 2024 rules, companies need not treat data as important unless regulators have notified them or published it as such, which brought welcome clarity.

Operators of critical information infrastructure must store personal information and important data collected in China within China. Even companies not caught by that rule often host China data locally, because it simplifies the transfer analysis.

The 2024 rules also let pilot free trade zones publish their own “negative lists” of data that needs a formal transfer mechanism. Companies registered in a zone can export data that is not on its list without a security assessment, standard contract or certification, so check whether your entity sits in one.

Generative AI and algorithm rules

China regulates AI through a series of targeted rules rather than one law. These include rules on recommendation algorithms (2022), on “deep synthesis” such as deepfakes (2023) and the Interim Measures for Generative AI Services, in force since 15 August 2023.

The Interim Measures apply to generative AI services offered to the public in China. Internal research and development that is not offered to the public falls outside them. Providers must use training data from lawful sources, respect intellectual property, obtain consent where personal information is used and govern content in line with Chinese law.

Services that can influence public opinion must complete algorithm filing and a security assessment before launch. In broad terms, a provider developing or deploying its own model goes through filing, while an app built on an already-filed model goes through a lighter local registration. Either way, it takes time.

Labelling AI-generated content

Since 1 September 2025, the AI content labelling measures and a mandatory national standard require two kinds of label. Explicit labels are visible to users, such as a text notice or watermark. Implicit labels are hidden in the file’s metadata so that machines can detect them.

The duties fall on AI service providers and on platforms that distribute content. Platforms must check for implicit labels, add visible labels where content is AI-generated or suspected to be, and give users tools to declare AI content they upload.

China is not alone here. The EU AI Act, India’s amended IT Rules and South Korea’s AI Basic Act all push in a similar direction, so labelling built for China can often be reused elsewhere.

How global companies usually approach China

The first decision is scope. Some companies choose not to serve China and enforce that through their terms, payments and app store settings. Others launch fully, and most choose something in between.

Companies that launch typically separate China operations: local hosting, a distinct data flow design and, for AI services, a local entity or partner to manage filing and content governance. A global model or feature usually cannot simply be switched on for users in China.

Whatever you choose, write the decision down and make the product match it. Regulators and partners will look at what your systems actually do, not only at what your policies say.