--- title: "Before a Chinese AI model goes public, it has to file a six-part dossier" date: 2026-10-02 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-generative-ai-filing-system language: en --- # Before a Chinese AI model goes public, it has to file a six-part dossier > China runs a two-track gate for generative AI (生成式人工智能): a large-model filing (大模型备案) for model builders and a lighter registration (登记) for apps that merely call an already-filed model. As of 30 April 2026, 868 services had filed and 530 apps had registered. A Chinese AI product does not launch when the model is good. It launches when the paperwork clears. Before a generative-AI (生成式人工智能) service can face the public in China, its provider must pass a gate that has no direct equivalent in the US or EU: file the model, pass a security assessment, and keep a dossier regulators can inspect. The system is older than the headline labeling rules, and it is the reason most Chinese AI apps carry a filing or registration number somewhere in their fine print. ## The two tracks The regime has two lanes: - **Large-model filing (大模型备案).** A provider that trains or substantially fine-tunes its own model must file with the Cyberspace Administration of China (CAC, 国家网信办) and pass a security assessment. This is the track for the model builders. - **Registration (登记).** A provider that only calls an already-filed model through an API — building an app on top — completes a lighter registration with its provincial cyberspace office. Same model, far less paperwork. Both sit under the Interim Measures for the Management of Generative AI Services (《生成式人工智能服务管理暂行办法》), in force since 15 August 2023. A separate algorithm filing (算法备案), dating to the 2022 Algorithmic Recommendation Provisions and the 2023 Deep Synthesis Provisions, covers the recommendation and synthesis mechanics. ## What the filing dossier contains For a full large-model filing, providers submit a six-part package: - An application form describing the model, its functions, audience and service scope. - A security self-assessment report. - A model service agreement (covering content safety and user responsibility). - Corpus annotation rules. - A keyword filter list. - An evaluation question set — plus API access or test accounts so reviewers can probe the model. The reference yardstick is the technical standard TC260-003, which expects lawful training data, documented annotation, and a high safety-pass rate on model outputs. ## The numbers behind the gate The CAC publishes periodic bulletins. In its 2026 update covering March–April, it reported: - 72 new services filed with the national CAC in those two months. - 49 new applications registered with provincial offices. - **Cumulative totals as of 30 April 2026: 868 services filed, 530 applications or functions registered.** Subsequent bulletins have kept the count climbing. Tracking analyses that compile the CAC notices put the totals higher by mid-2026, but the 868 / 530 figure is the last officially announced cumulative snapshot we can cite directly. ## What happens if you skip it The consequences are written into the rules. Under the Algorithmic Recommendation Provisions, a provider that fails to file can receive a warning, a public reprimand, and an order to rectify. If it refuses or the violation is serious, regulators may suspend information updates and impose a fine of RMB 10,000 to 100,000 — roughly US$1,400 to US$14,000 at prevailing exchange rates. Enforcement has been visible: local CAC offices have summoned and penalized AI apps that launched without filing or registration, and app stores have been told to remove non-compliant face-swap and similar tools. ## Why a two-track system The split is pragmatic. A base-model lab controls training, safety evaluation and updates — it carries the heavier duty. A downstream app may not be able to see or change the model's internals; it controls the interface, retrieval and prompts. Requiring it to re-file the whole model would be redundant, so registration recognizes its narrower role. The effect is a two-layer market: a growing set of compliant base models, and a faster-growing set of applications built on top of them. That favors vendors — cloud providers, model companies — that can turn one model into infrastructure for many products. ## What it does not do Filing is a market gate, not a quality verdict. A filed model is not certified accurate, safe in every context, or commercially successful. The dossier reveals little about revenue, retention or inference volume. And the system is inward-facing: it governs services offered to the public inside China, with weaker reach across borders. ## Honest limitations The two-track structure (large-model filing + registration), the legal bases (2023 Interim Measures art. 17; 2022 Algorithmic Recommendation Provisions; 2023 Deep Synthesis Provisions), the six-part dossier, and the TC260-003 yardstick are drawn from legal analyses (ICLG) and the comparative-ai.org China tracker. The cumulative counts — 72 filed and 49 registered in March–April 2026; 868 filed and 530 registered as of 30 April 2026 — are from the CAC's own bulletin as reprinted by Chinese media (China Star Market, regional CAC offices) and corroborated by third-party trackers. The RMB 10,000–100,000 fine and its US$1,400–14,000 equivalent are quoted from ICLG's analysis of the Algorithmic Recommendation Provisions. We have not independently audited any single filing, and later bulletins may show higher totals than the 30 April 2026 snapshot. No other RMB amounts appear. Analysis is current to 2 October 2026. ## What readers can do now - **If you launch an AI app in China**, decide early whether you are filing a model or registering an app — calling an already-filed model via API puts you on the lighter registration track. - **If you build on Chinese models**, expect to display the base model's filing or registration number in your product; pick providers with clean compliance records. - **If you compare regimes**, note China's ex-ante gate (no filing, no public service) has no direct US or EU equivalent — it is a market-entry control, not just a post-hoc audit. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-generative-ai-filing-system Free to quote with attribution and a link to the original.