Under the Provisions on the Administration of Algorithm-Generated Recommendation Services and the Provisions on the Administration of Deep Synthesis Services, an algorithm filing requires entity information, algorithm information, a safety assessment and undertakings. By 2024 the filing system was operating routinely. Whether materials are complete and consistent determines whether a submission passes at the first attempt, rather than whether gaps can be filled afterwards.
Entity information tests accountability
Entity materials usually include the business licence, legal representative details, a contact person and contact details, and the algorithm safety officer. The review is less about whether the file set is complete than about whether the entity can be reached and responsibility can be attached to a named person. Outdated contact details, or a safety officer who does not match the actual role, are common reasons for rejection. Where several entities cooperate on a project, their respective roles must be explained, so that the situation does not arise where the filing entity does not operate the service and the operator has no filing.
Algorithm information must explain the mechanism
Algorithm information covers the name, type, application scenario, products launched and an explanation of how the mechanism works. The difficulty lies in translating technical logic into language a reviewer can follow. A clear account of inputs, processing and outputs, and of how recommendation or generation affects users, works better than stacked terminology. Where the algorithm has recently changed in substance, the materials should describe the current version rather than reuse an older description.
Assessment, undertakings and common corrections
The safety assessment explains risk identification, control measures and residual risk; the undertakings confirm compliance duties. The two often drift apart: the assessment describes one set of measures while the undertaking is a generic template. Reviewers focus on whether measures are specific and executable. Writing the key measures from the assessment into the undertakings and internal policies, item by item, visibly reduces corrections.
Four defects are worth listing separately. Inconsistency across materials, such as a product name or algorithm name differing between a table and a narrative. Mechanism descriptions so general that the service form cannot be determined. Safety measures stated as principles, without an owner or a frequency of execution. Missing attachments or outdated versions. A correction request is not a failure, but each round extends the launch timeline and disrupts product scheduling.
Change, cancellation and interaction with generative AI filing
Filing is not a one-off act. Changes to entity information, algorithm function or launch channel require a change filing; termination of service requires cancellation. Generative AI services with public-opinion attributes or social-mobilisation capacity also require generative AI service filing, which focuses on the service and its safety assessment rather than on the algorithm mechanism. Where a product has both recommendation and generation functions, the two sets of materials should sit in one register with a consistent external description.
A self-check order before submission
First confirm that entity and product information match across all documents. Then check that the algorithm description maps to specific functions. Next, verify that the measures in the safety assessment appear in the internal policies. Finally, confirm attachment versions and signatures. The value of this order is that structural problems are resolved before wording is polished.
For material preparation and responding to corrections, see our algorithm filing service; for synthetic-content labelling, see AIGC labelling; for training-data arrangements, see data compliance.
This article is based on the Provisions on the Administration of Algorithm-Generated Recommendation Services, the Provisions on the Administration of Deep Synthesis Services and public information about the algorithm filing system in 2024. It is for reference only and does not constitute legal advice.