Provisions on the Administration of Deep Synthesis in Internet Information Services

Assume primary responsibility for information security by establishing management systems for user registration, review of algorithm mechanisms, science and technology ethics review, content-publication review, data security, personal information protection, anti-telecom-fraud and emergency response, backed by secure and controllable technical safeguards. Also adopt and publish management rules and a platform convention, refine the service agreement, and prominently remind technical supporters and users of their information-security duties.

Applies when

Deep synthesis service providers. The ethics-review limb connects directly to the ethics committee required by the Science and Technology Ethics Review Measures and should be built once, not twice.

Evidence to retain

• · The eight management policies with their issuance and effective-date records
• · Establishment documents and review records for the science and technology ethics committee, with a note on the joint build with the Ethics Review Measures
• · Screenshots of the published management rules and platform convention
• · Information-security notice clauses in the service agreement and evidence of how they are displayed
• · Description of the technical safeguards, with a security and controllability justification

Estimated first-time effort: 20–45 person-days (estimate, not a commitment)

Authenticate users’ real identity through their mobile number, identity-document number, unified social credit code or the national network identity authentication public service, and do not provide information-publishing services to users who have not been authenticated. Where an enterprise customer is authenticated by its unified social credit code, the individual operating accounts underneath must still be traceable.

Applies when

Products offering digital humans, face swapping, voice cloning or AI writing with a publishing capability; purely local-inference tools without publishing should still be assessed against the “information-publishing service” concept.

Evidence to retain

• · Real-identity authentication design: factors used, verification channel and failure handling
• · Product evidence that publishing is blocked for unauthenticated users
• · Mapping and audit records linking enterprise customer accounts to the individuals operating them
• · Note on data minimisation and encrypted storage of authentication data

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

Strengthen content management by reviewing users’ input data and synthetic outputs, whether by technical means or manually. Maintain a rumour-refutation mechanism: on finding that the service has been used to create, copy, publish or spread false information, take refutation measures promptly, retain the records, and report to the cyberspace authority and other competent departments. Also provide convenient entry points for user appeals and public complaints and reports, publishing the handling process and response times.

Applies when

All deep synthesis service providers. Input review for face-swapping, digital-human and voice-cloning products must additionally verify authorisation for the use of a person’s likeness and voice.

Evidence to retain

• · Input and output review policy covering technical rules and the manual escalation path
• · Verification workflow for likeness and voice authorisation, with sample authorisations
• · Description of the rumour-refutation mechanism, refutation records and receipts for reports to the authority
• · Screenshots of the appeal and complaint entry points and the published process and response times
• · Retention policy for review records

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

Deep synthesis service providers and technical supporters must strengthen training-data management and take the necessary measures to keep training data secure; where training data contains personal information, the personal-information rules apply. They must also strengthen technical management by periodically reviewing, evaluating and validating the mechanisms of their generative and synthetic algorithms. Both duties bind technical supporters, that is base-model and SDK vendors.

Applies when

Both providers and technical supporters. The cadence, method and conclusions of the periodic review must be documented; inspections ask for the successive records, not just the policy.

Evidence to retain

• · Training-data security policy with grading and access-control records
• · Lawful basis and processing records for personal information within training data
• · Plan for periodic algorithm review, evaluation and validation, stating cadence, metrics and method
• · Successive review reports and the remediation records for findings
• · Agreement allocating duties between the technical supporter and the provider

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

Art. 16: apply technical measures to add a label that does not impede use to any content generated or edited through the service, and retain log information as required by law and national rules. Art. 17: where the service provides intelligent dialogue or intelligent writing that simulates a natural person generating or editing text; synthetic or imitative voice generation or editing that significantly alters personal identity characteristics; face generation, face swapping, face manipulation or pose manipulation in images and video, or editing that significantly alters personal identity characteristics; immersive realistic scene generation or editing; or any other function that generates or significantly alters content — and the output may cause the public to confuse or mistake it — a conspicuous label must be placed at a reasonable position or area within the content. For deep synthesis services outside that list, the provider must offer a conspicuous-labelling function and remind users that they may use it.

Applies when

All deep synthesis service providers. The five categories in Art. 17 §1 must be conspicuously labelled, while everything else must merely offer a labelling function and prompt users to use it; the two require quite different product implementations and must be separated at the requirements stage.

Evidence to retain

• · Mapping table comparing each service function against the five categories in Art. 17 §1
• · Implementation note for conspicuous and implicit labels, with samples for each modality
• · Design justification that the label does not impede use
• · Log retention policy and screenshots of the retention configuration
• · Description of the mechanism for detecting and responding to deleted, altered or concealed labels

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

A deep synthesis service provider with public-opinion attributes or social-mobilisation capability must complete algorithm filing, and amendment and cancellation filings, under the Algorithmic Recommendation Provisions. Where it develops and launches a new product, application or feature with such attributes, it must carry out a security assessment under the applicable national rules. The assessment is therefore not a one-off but recurs with each significant launch.

Applies when

Deep synthesis providers with public-opinion attributes or social-mobilisation capability. Product teams should add a “does this trigger a security assessment?” gate to the release checklist, or a feature launch itself becomes the breach.

Evidence to retain

• · Algorithm filing pack and a screenshot of the published filing number
• · Records of amendment and cancellation filings
• · Trigger-determination table and release checklist for new products, applications and features
• · Successive security assessment reports and submission records

Estimated first-time effort: 15–40 person-days (estimate, not a commitment)