AI Answer Library
Short answer
First determine whether you are using AI internally or offering a generative AI service to the public in mainland China — that single distinction sets the scale of your compliance burden. China's Interim Measures for the Administration of Generative AI Services primarily bind parties that provide such services to the public; enterprises that develop and apply the technology without offering a public-facing service fall outside their scope. The Data Security Law, the Personal Information Protection Law and the cybersecurity classified protection scheme, by contrast, apply to everyone. In practice, classify your data first, decide whether it may leave your perimeter second, and only then discuss model selection.
The table maps the instruments a mainland-China enterprise AI project most often touches, split into what each governs, who it binds, and the practical implication — enough to judge at kickoff which ones concern you. Two caveats: this is directional only, cites no article numbers, and does not substitute for legal advice. Sector-specific rules commonly layer on top in healthcare, finance, education and industry, and regulatory interpretation evolves with practice. The binding conclusion must come from your own counsel, applied to your specific business.
| Instrument / requirement | What it governs | Who it binds | Practical implication |
|---|---|---|---|
| Interim Measures for the Administration of Generative AI Services | The act of providing a generative AI service: content safety, lawful training data, labelling, user management | Parties offering generative AI services to the public in mainland China; development and application without a public offering is out of scope | Decide which category you are in first. If you will serve the public, put algorithm filing and security assessment into the project plan |
| Data Security Law | Classified and tiered data protection, and the handling obligations attached to important data | Every organisation carrying out data processing activities within mainland China | Grade the corpus first, then decide what may enter a knowledge base and whether it may leave the network |
| Personal Information Protection Law | Notice and consent, minimisation, separate consent for sensitive data, and responding to individual rights requests | Every organisation processing personal information, internal systems included | Redact or obtain authorisation before ingesting customer records, and keep deletion and export technically feasible |
| Cybersecurity classified protection (MLPS 2.0) | Security controls and management measures by system grade, plus assessment | Network operators in mainland China; the grade follows system criticality and impact | An AI system is usually graded together with the business system it serves — budget the assessment cycle into the schedule |
| Cross-border data transfer rules | The routes and preconditions for sending data abroad: security assessment, standard contract, or certification | Any processor that needs to transfer data outside mainland China | Calling an overseas model API can itself constitute an outbound transfer — assess it separately rather than assuming it is fine |
| Labelling of AI-generated content | AI-generated or synthesised content shown to the public must be identifiable | Products that present generated content directly to external users | Build labelling in at design time; retrofitting after launch is expensive |
Question one: will this system offer a generative AI service to the general public in mainland China? If yes, the compliance path becomes materially heavier — algorithm filing, content safety mechanisms and user management all need advance planning, with legal counsel engaged at design time. If no (internal staff only, or contracted customers in a controlled environment), the centre of gravity shifts to data and security. Question two: what data will it touch? Following the Data Security Law's tiering logic, sort the corpus into three bands — publicly available, internal-general, and sensitive or personal — and state which systems each band may enter. Question three: will data leave the internal network, or the country? Calling an overseas model API may constitute an outbound transfer, and the answer here often decides self-hosting versus API on its own. Question four: what protection grade applies and when is the assessment? Put the assessment cycle in the schedule rather than the week before launch. Walking these four in order resolves most disputes while the project is still on paper.
Compliance is not only paperwork; it imposes hard requirements on the system, and without them you cannot demonstrate compliance at all. First, complete audit logs: who asked what and when, which documents were retrieved, what the model returned — all traceable and retained for the required period. Second, permission isolation: retrieval must filter by user and department, or anyone can extract material they are not cleared to see simply by asking. Third, deletable and exportable data: individuals may demand erasure of their personal information, and if the knowledge base cannot locate and remove one person's data, that obligation cannot be met. Fourth, filtering and refusal: requests outside the authorised scope must be declined rather than answered anyway. Fifth, the ability to switch things off: take a capability or a corpus offline quickly when something goes wrong. All five are straightforward under self-hosting; on an architecture built directly on external APIs they must be added at the gateway layer.
Where this applies
People also ask
Self-hosted LLM or public API — how do I choose?
How much does it cost for a company to self-host a large language model?
Should a company buy software outright or subscribe to SaaS?
Ygg Lab Token — API Quota Gateway
Every model call, clear and in control — nine upstream provider types behind a single sk- key, with quota, logs and billing you own.
Township Health Agent
Policy Q&A, health records, family-doctor sign-up and check-up booking, all inside the county hospital WeChat account — nothing for residents to install.