Interim Measures for the Management of Generative AI Services

Conduct pre-training, fine-tuning and other training-data processing lawfully: use data and foundation models from lawful sources; do not infringe others’ intellectual property rights; where personal information is involved, obtain the individual’s consent or rely on another statutory basis; take effective measures to improve training-data quality, strengthening its authenticity, accuracy, objectivity and diversity; and comply with the other requirements of the Cybersecurity Law, the Data Security Law, the Personal Information Protection Law and the supervisory requirements of the competent authorities.

Applies when

Every generative service provider that builds or fine-tunes a model, including teams retraining third-party open-source foundation models — the lawfulness of the foundation model’s own provenance is itself an inspection item.

Evidence to retain

• · Training-data provenance register covering dataset name, source, acquisition method, licence or authorisation, and usage restrictions
• · Foundation-model licences and a compliance note comparing open-source licence terms against actual use
• · IP risk screening records (copyright, trade marks, database rights) and the removal list
• · Statement of the lawful basis for personal information in the corpus, with de-identification and cleaning records
• · Training-data quality evaluation report across authenticity, accuracy, objectivity and diversity

Estimated first-time effort: 20–50 person-days (estimate, not a commitment)

Where data is annotated during technical development, adopt clear, specific and operable annotation rules that meet these Measures; run annotation quality evaluation with sampling checks on accuracy; and provide the necessary training to annotators to raise their awareness of legal compliance, supervising and guiding them in performing the work properly. Outsourcing annotation does not transfer these duties away from the provider.

Applies when

Every model team doing SFT, RLHF, safety alignment or domain annotation, including where annotation is outsourced to onshore or offshore vendors.

Evidence to retain

• · Annotation rulebook, including criteria for unlawful and harmful information and rules for handling discriminatory content
• · Sampling plan for annotation quality and successive evaluation reports, showing sampling rate, accuracy and rework records
• · Annotator training materials, attendance records and test results
• · Vendor qualification review, contract and oversight records for outsourced annotation

Estimated first-time effort: 10–25 person-days (estimate, not a commitment)

Assume the responsibilities of a network information content producer and the duties of network information security, and where personal information is involved, the responsibilities of a personal information processor. Execute a service agreement with registered users that sets out both parties’ rights and obligations; state and publish the intended user groups, settings and uses of the service, guiding users towards an informed and lawful understanding of generative AI; and adopt effective measures to prevent minors from becoming over-reliant on, or addicted to, the service.

Applies when

Every provider offering generative services to the public in China. Products in education, companionship, mental-health and entertainment settings need anti-addiction measures built as evidenced product features, not merely agreement clauses.

Evidence to retain

• · User service agreement covering rights and obligations, usage limits and notice of labelling duties
• · Screenshots of the published statement of intended user groups, settings and uses
• · Design documentation and release evidence for anti-addiction measures (time limits, content grading, guardian controls)
• · Captured screens for usage guidance and risk warnings

Estimated first-time effort: 8–20 person-days (estimate, not a commitment)

Protect users’ inputs and usage records as required by law: do not collect unnecessary personal information; do not unlawfully retain inputs and usage records capable of identifying the user; and do not unlawfully provide users’ inputs and usage records to others. Promptly handle individuals’ requests to access, copy, correct, supplement or delete their personal information. This directly constrains the practice of feeding user conversations into training sets by default, which needs separate, revocable authorisation and a deletion path that actually works.

Applies when

All conversational and generative products, including on-premise deployments that still send telemetry back to the provider.

Evidence to retain

• · Field inventory, retention periods and minimisation justification for inputs and usage records
• · Product evidence of a separate, revocable authorisation for using conversations in training
• · De-identification design for identity-revealing content and a technical verification report
• · End-to-end records showing rights requests executed across conversation logs, training sets and vector stores

Estimated first-time effort: 10–25 person-days (estimate, not a commitment)

Providers must label generated images, video and similar content as required by the Deep Synthesis Provisions. From 1 September 2025 the method is set by the AI Content Labelling Measures and mandatory national standard GB 45438-2025: content falling within Art. 17 §1 of the Deep Synthesis Provisions requires an explicit label, and all generated and synthetic content requires an implicit label in the file metadata.

Applies when

Every generative service provider, including API-only model platforms — whether downstream applications can label correctly depends on whether the upstream API carries the metadata in its output.

Applies from: 2025-09-01

Evidence to retain

• · Labelling design note covering explicit label placement and style and the implicit metadata fields
• · API documentation showing metadata carried in outputs, plus captured traffic or sample files
• · Labelled samples retained for each modality: text, image, audio, video and virtual scenes
• · Self-assessment records against GB 45438-2025

Estimated first-time effort: 10–30 person-days (estimate, not a commitment)

On discovering unlawful content, promptly stop generation, stop transmission and remove it, remediate through measures such as model optimisation and retraining, and report to the competent authority. On discovering a user engaging in unlawful activity through the service, issue warnings, restrict features, or suspend or terminate the service under law and contract, retain the relevant records, and report to the competent authority. Also maintain a complaints and reporting mechanism with a convenient entry point, publish the handling process and response times, and deal with and respond to reports promptly.

Applies when

Every generative service provider. The requirement to remediate through “model optimisation and retraining” signals that regulators expect root-cause fixes, not just filter patches.

Evidence to retain

• · Content-safety review policy and the criteria for identifying unlawful content
• · Ticket log for unlawful content, recording detection time, action taken and completion time
• · Model remediation records: alignment data, safety fine-tuning, and before-and-after evaluation results
• · Records of reports to the competent authority and the receipts
• · Screenshots of the complaints entry point, the published process and response times, and the disposition log

Estimated first-time effort: 15–35 person-days (estimate, not a commitment)

A provider of generative AI services with public-opinion attributes or social-mobilisation capability must carry out a security assessment under the applicable national rules and complete algorithm filing, together with amendment and cancellation filings, under the Algorithmic Recommendation Provisions. Art. 24 of those Provisions sets the filing deadline at ten working days from the start of the service, and once filed the provider must display the filing number prominently on its website or application with a link to the published information. From 1 September 2025, filing and assessment submissions must also include the labelling materials required by Art. 12 of the Labelling Measures.

Applies when

Public-facing conversational products, content-generation platforms and AI applications with community or distribution features are almost always treated as having public-opinion attributes or social-mobilisation capability. Purely internal enterprise deployments that are not offered to the public typically do not trigger it, but opening a public trial entry point can change the analysis.

Deadline: Algorithm filing: within ten working days from the start of the service (Art. 24 of the Algorithmic Recommendation Provisions).

Evidence to retain

• · Determination record and reasoning on public-opinion attributes or social-mobilisation capability
• · Security assessment report with submission and feedback records
• · Algorithm filing pack: provider name, service form, application field, algorithm type, algorithm self-assessment report and the content to be published
• · Screenshots showing the filing number and the link to published information in a prominent product location
• · Records of amendment and cancellation filings
• · The labelling documentation included in the filing and assessment submissions

Estimated first-time effort: 20–50 person-days (estimate, not a commitment)