Your team wants to fine-tune a model on a Chinese-language corpus, or your Chinese subsidiary holds two years of customer conversations that would make an excellent training set. Someone in legal asks whether PIPL applies, and the honest answer is that it probably does, that consent collected for customer service does not cover model training, and that the data may not lawfully leave the country in the form you planned. This guide sets out how the Personal Information Protection Law governs training data, where the bottlenecks actually are, and what a defensible file looks like.
Who this guide is for
- You are building or fine-tuning a model using data that includes information about people in China, whether you collect it yourself or license it.
- You are buying a Chinese dataset and need to know what the seller’s consent and provenance records actually support.
- You already operate in China and must decide whether training data can move to a parent model or a third-country vendor.
Stage: Execute. You are past the question of whether China matters. The work now is evidence: lawful basis, records, assessments and a transfer route that survives review.
Does PIPL reach your training pipeline?
PIPL applies to processing carried out inside China. It also applies outside China where the purpose is to provide products or services to individuals in China, or to analyse and assess the behaviour of individuals in China. Training is squarely within scope, because “processing” covers collection, storage, use, processing, transmission, provision, disclosure and deletion.
Two consequences follow. A model trained abroad on data about Chinese users can be caught even if no data is stored in China, and an overseas processor in scope must appoint a local representative or designated entity and record it with the regulator. “We train offshore” is not a defence.
The scope turns on what counts as personal information: information relating to an identified or identifiable natural person, excluding anonymised information. De-identified data remains personal information, because it can still be re-linked with additional information. Only anonymisation, where re-identification is impossible and irreversible, takes data outside the law, and the standard is demanding.
Legal bases, and why consent is usually the wrong default
| Legal basis | Fit for AI training | Practical note |
|---|---|---|
| Consent | Workable for a purpose-built dataset | Must be voluntary, informed, specific and withdrawable |
| Necessary to conclude or perform a contract with the individual | Covers service delivery, not model training | Stretching it to training is the common error |
| Necessary to perform a statutory duty or obligation | Narrow | Rarely available to a commercial model developer |
| Public interest, public health emergency, or protection of vital interests | Very narrow | Not a general research exemption |
| Personal information the individual has himself disclosed or that is otherwise lawfully disclosed | Closest thing to a research basis | Limited to a reasonable scope and the individual’s reasonable expectations |
| Other circumstances provided by law | Reserved | Watch for sector rules |
The decisive point is what is missing. PIPL has no legitimate interests basis, which is the single biggest difference from GDPR and breaks most training data plans drafted in Europe. If training cannot fit one of the listed bases, you need consent, and consent must be specific to training rather than bundled into general terms.
Separate consent in practice
Separate consent (单独同意) is a distinct, granular consent obtained for one specific act. It is required when providing personal information to another processor, when publicly disclosing it, when processing sensitive personal information, when transferring it abroad, and for automated decision-making that pushes information or marketing to individuals.
Applied to training data, four problems recur. A privacy policy checkbox is not separate consent. Consent for “service improvement” does not cover training a foundation model for other customers. Withdrawal must be as easy as giving consent, which means you need a way to remove data from a training set or to show that a trained model no longer contains it. And retrospective consent for scraped or legacy data is usually impractical, which is why experienced teams redesign the dataset instead of the consent flow.
Sensitive data, minimisation and lawful source
Sensitive personal information
Sensitive personal information includes biometric characteristics, religious belief, specific identity, medical health, financial accounts, whereabouts and trajectory, and information about minors under 14. AI uses that commonly trigger it include face and voice data for authentication or emotion analysis, health data in medical models, precise location in mobility and delivery products, and any data drawn from children’s services.
Processing it requires a specific purpose and sufficient necessity, strict protection measures, separate consent, and notice of the necessity and the impact on the individual’s rights. An impact assessment is mandatory. For biometric training data, the practical test is whether a less intrusive alternative, such as synthetic or aggregated data, would achieve the same result.
Minimisation and provenance
PIPL requires a clear and reasonable purpose, processing directly related to it, and the minimum scope necessary. For training, that translates into defining the purpose before collection, separating training from service delivery in your notices, preferring aggregated or synthetic data, and keeping a provenance record for every corpus.
Publicly available does not mean free to use. The basis for processing publicly disclosed personal information is limited to a reasonable scope and to what the individual would reasonably expect, and scraping at scale can also engage the Cybersecurity Law, unfair competition rules and the Network Data Security Management Regulations in force since 2025. A dataset bought from a broker needs contractual representations plus evidence of the original consent, not a warranty alone.
Moving training data out of China: three routes
| Route | When it applies | Mechanism | Typical timing |
|---|---|---|---|
| Security assessment | Critical information infrastructure operators exporting any personal information; any processor exporting important data; export of more than 1,000,000 individuals’ non-sensitive personal information or more than 10,000 individuals’ sensitive personal information in a year | Application through the provincial cyberspace administration, assessment by the national administration | Commonly several months |
| Standard contract | Between 100,000 and 1,000,000 individuals’ non-sensitive personal information, or fewer than 10,000 individuals’ sensitive personal information | Use the prescribed template and file it with the provincial administration | Weeks to a couple of months |
| Certification | Voluntary route, often used for intra-group transfers | Certification by an accredited body | Weeks to months |
| Exemption | Fewer than 100,000 individuals’ non-sensitive personal information in a year | No mechanism required, but lawful basis, notice and separate consent still apply | Not applicable |
Thresholds are cumulative from 1 January of the current year and counted by individual rather than by record, so a single user with a thousand messages counts once. Confirm the current text and any free trade zone negative list before relying on a threshold. These rules continue to evolve, and this article reflects the position as of late 2026.
Impact assessment and the cost of getting it wrong
The PIPIA
A personal information protection impact assessment (PIPIA) is triggered by processing sensitive personal information, automated decision-making, entrusting processing, providing information to third parties, public disclosure, cross-border transfer and other processing with a major impact on individuals. It must cover the purpose and method, the impact on rights and interests, security risks, protective measures and a conclusion, and the report and processing records must be kept for at least three years.
For an AI company the assessment should reach beyond the statute’s checklist: where each corpus came from, how labelling was done, whether the model can memorise and regurgitate personal information, how outputs are filtered, and whether evaluation sets create re-identification risk.
Penalties
The administrative ladder starts with an order to correct, a warning and confiscation of illegal gains. Refusal to correct can attract a fine of up to RMB 1 million. Serious violations can attract a fine of up to RMB 50 million or 5 percent of the previous year’s turnover, suspension of business, revocation of licences, and fines of up to RMB 1 million on directly responsible individuals, with a possible bar from management roles.
Three other exposures sit alongside the fine. Individuals can sue, procuratorates can bring public interest litigation, and unlawful sale or provision of personal information can be prosecuted as a crime. For a foreign company the practical sanction is often commercial: an app removed, a filing refused, or a Chinese partner terminating the contract to protect its own licence.
What to do next
- Inventory every dataset by source, legal basis and whether it contains sensitive or children’s data, and mark the ones with no defensible basis.
- Stop using service consent as training consent; build a separate consent flow with a working withdrawal path, or replace the data.
- Write a provenance record for each corpus, including source channel, licence terms, cleaning and de-identification steps.
- Run a PIPIA before training begins, not before launch, and refresh it when the corpus or the model purpose changes.
- Map every cross-border flow to one of the three routes and record the individual count from 1 January.
- Apply de-identification and synthetic data where biometric or health data would otherwise be necessary, and document why the residual data is needed.
- Give the board a written statement of the lawful basis for the training programme, because personal liability attaches to the people who approved it.
Frequently Asked Questions↑ Back to article
Does PIPL apply to training data we process outside China?
Yes, in two situations: where you process personal information outside China in order to provide products or services to individuals in China, and where you process it to analyse or assess the behaviour of individuals in China. A model trained abroad on data about Chinese users can therefore be caught. An overseas processor in scope must also appoint a local representative or designated entity and record it with the regulator.
Can we rely on legitimate interests to train a model?
No. PIPL does not include legitimate interests among its legal bases, which is the single biggest difference from GDPR. Training must fit consent, necessity for a contract with the individual, a statutory duty, the narrow public-interest grounds, or processing of information the individual has himself disclosed or that is otherwise lawfully disclosed within a reasonable scope. If none fits, you need consent.
Is de-identified training data still personal information?
Yes. De-identification reduces risk and is encouraged, but the data remains personal information because it can still be linked back to an individual with additional information. Only anonymisation, where re-identification is impossible and the process cannot be reversed, takes data outside PIPL. Keep the mapping tables and re-identification keys under strict access control.
Do we need separate consent to use customer conversations for training?
Yes. Consent collected to deliver a service does not extend to training a model, and a general privacy policy acceptance is not separate consent. You need a distinct, granular consent for the training purpose, or another legal basis, together with notice of what the data will be used for and how withdrawal works. Using legacy conversations for training is one of the most common exposure points we see.
How can we move training data out of China?
There are three routes: a security assessment for critical information infrastructure operators and for volumes above the published thresholds, a standard contract filed with the provincial cyberspace administration, or personal information protection certification. Export of fewer than 100,000 individuals' non-sensitive personal information in a year is exempt from the mechanisms, but notice, a lawful basis and separate consent still apply. Thresholds are cumulative from 1 January and counted by individual.
What must a personal information protection impact assessment cover?
The assessment must describe the processing purpose and method, the impact on individuals' rights and interests, the security risks, the protective measures taken and the conclusion. It is mandatory for sensitive personal information, automated decision-making, entrusting processing, provision to third parties, public disclosure and cross-border transfer. The report and processing records must be retained for at least three years.
Who can be held personally liable for a PIPL breach?
For serious violations, the regulator can fine the directly responsible individuals up to RMB 1 million and bar them from senior management or personal information protection roles for a period. The company may also be fined up to RMB 50 million or 5 percent of the previous year's turnover, ordered to suspend service or have its licence revoked. Directors who approved a training data programme without a lawful basis are within the personal exposure.