Algorithm Registration and LLM Filing
Filing with the CAC is not a form. It is a technical dossier — mechanism description, security self-assessment, compliance systems — and it decides whether the product may operate at all.
Registration is the gate, not the paperwork after it
Since the Provisions on the Administration of Algorithm Recommendations for Internet Information Services took effect in 2022, algorithm registration moved from recommended practice to a statutory obligation. The Interim Measures for the Management of Generative AI Services then added a second layer: a generative service must complete algorithm registration and, where a large model is involved, large-model filing before it may lawfully operate.
Two consequences follow. First, the obligation attaches to the service, not to the entity’s intention — a product that is live without filing is not “exposed to a risk”, it is in a non-compliant state. Second, review has become technical: filings are examined against the actual architecture, so a document set written to satisfy a template will be returned.
What we do
- Scope analysis: which algorithms require registration, which require large-model filing, and which require both
- Mechanism documentation: algorithm principle description, data sources and processing, model and labelling system
- Security self-assessment: the core document — risk identification, mitigation measures, user-rights protection, emergency response
- Compliance systems: content review, complaint handling, ethics review, and the internal records that evidence them
- Submission and follow-up: initial review, correction rounds, re-review, publication and registration number
- Parallel filing: coordinating algorithm registration with large-model filing so the two dossiers do not contradict each other
What is different for an AI company
- The filing entity must match the operating and collecting entity. Model trained by A, charged for by B, licensed under C — the filing names one entity and the business another, and that is the first question in an inspection.
- Filing is not one-off. Model version upgrades, changes in service form, or a broader audience can each trigger a change filing. Keep a version log and a change register as routine rather than reconstructing materials later.
- Security assessment is a second gate. Services with public-opinion attributes or social-mobilisation capacity need a security assessment in addition to filing; filing alone leaves a procedural gap.
- Entity eligibility under foreign or VIE structures must be confirmed early. Filing imposes requirements on entity nature and shareholding; offshore control and round-trip investment affect acceptance.
- The data-source description in the filing must survive later audit. Provenance, annotation method and cleaning process stated at filing will be asked for as records in inspections and diligence — write it to a verifiable standard from the start.
Typical situations
Situation 1: The assessment is returned, and the clock restarts
The security self-assessment reads as boilerplate: the mechanism description restates marketing material, the risk analysis names no concrete risk, and data sources are described as “publicly available” without stating how. Correction rounds typically cost two to four weeks each, and a returned filing usually has more than one defect.
Our approach: write the assessment against the architecture. Data provenance is traced to source categories, the risk section names the specific failure modes the product can produce, and mitigation is described in terms of mechanisms that exist in the system rather than policies that exist in a document.
Situation 2: The product is already live, and filing becomes remediation
A service launches first and files later. When filing is required under enforcement pressure, the timeline is no longer the firm’s to control, and the available remedies narrow — rectification within a deadline, removal from app stores, or administrative penalty.
Our approach: sequence filing before launch where the service is a recommendation or generative service, and where the product is already live, run a gap assessment first so that remediation addresses the actual exposure rather than only the document set.
Situation 3: Algorithm registration and large-model filing are treated as one thing
A team provides a generative service on top of a third-party or self-developed large model. Registration covers the algorithm; filing covers the model. Both apply. The documents overlap but the questions are different, and filing one without planning the other produces inconsistencies that surface at review.
Our approach: map the two obligations onto a single dossier plan, assign one owner, and write the overlapping sections once so that the description of data sources, training process and safety measures is identical across both.
Process and deliverables
| Stage | Work | Deliverables |
|---|---|---|
| 1 | Architecture and scope review | Filing map (what files, where, in what order) |
| 2 | Mechanism documentation | Algorithm principle description, data source statement |
| 3 | Security self-assessment and compliance systems | Assessment report + internal policy set |
| 4 | Submission, correction rounds, publication | Registration number + published disclosure text |
| Ongoing | Change filing on material product changes | Updated dossier + change log |
AI-specific deliverables
- Filing-scope map: which algorithms, models and service forms require filing or registration, and which entity files.
- Mechanism description pack: the technical write-up, safety-assessment inputs and compliance-policy set required at filing.
Frequently asked
How long does registration take? Two to three months from kick-off to registration number where the dossier is complete. A correction round adds two to four weeks. The variable is not the authority’s speed but the quality of the first submission.
Can we launch while the filing is pending? For recommendation and generative services, the obligation is to file before providing the service. Launching first and filing later is a compliance exposure that is difficult to defend during an enforcement campaign, and it changes the nature of the filing from compliance step to remediation.
We are a platform calling model APIs — who files? Filing obligations attach according to the role actually played: providing the service to the public, providing technical support, or both. A platform that re-presents model capability under its own brand is generally closer to a service provider than a pure conduit. The answer depends on the architecture, and it is worth resolving before launch rather than after.
What changes require a change filing? Material changes to the algorithm, its application scenario, or the model behind the service. Cosmetic iteration does not. The practical rule is to keep a change log and assess each release against the published disclosure, because the published text is what the filing is judged against.
Legal basis
- Provisions on the Administration of Algorithm Recommendations for Internet Information Services (effective 2022): registration obligation, self-assessment, disclosure
- Provisions on the Administration of Deep Synthesis of Internet Information Services (effective 2023): filing and labelling obligations for deep synthesis services
- Interim Measures for the Management of Generative AI Services: filing for generative services, training-data requirements, content obligations
- Data Security Law and Personal Information Protection Law: data provenance and personal information handling within training and inference
- Cybersecurity Law: network operation and content management obligations that attach alongside filing