--- title: "China's informatics plan pulls AI agents and on-device models into the filing system" date: 2026-11-02 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-informatics-plan-ai-filing language: en --- # China's informatics plan pulls AI agents and on-device models into the filing system > The National Informatisation Plan for 2026–2030 says agent, on-device model and anthropomorphic-interaction services will be brought into assessment and filing — but the filing mechanics are not written yet. A product team ships an AI assistant that runs on the phone rather than in a cloud. No server, no API key, no obvious "service provider" in the old sense. Six months ago that design felt like a way to stay outside the regulatory perimeter. A planning document released in Beijing suggests it will not stay that way. ## Key takeaways - The Central Cyberspace Affairs Commission issued the 《国家信息化"十五五"规划》 (National Informatisation Plan for the 15th Five-Year Plan period); the cyberspace regulator announced it on 23 October 2026. - One line in the security section matters more than the rest: 优化生成式人工智能备案机制, 建立重点领域应用安全风险评估制度, 完善智能体、端侧大模型、拟人化互动服务等新形态的评估备案 — optimise the generative AI filing mechanism and complete assessment-and-filing for new forms including agents (智能体), on-device large models (端侧大模型) and anthropomorphic interaction services (拟人化互动服务). - The plan is 印发 (issued) and operative as a plan. The filing expansion is a task statement, not a published procedure with forms, deadlines or thresholds. - What is already binding today sits elsewhere: the 《人工智能拟人化互动服务管理暂行办法》 took effect on 15 July 2026, and generative AI service filing has been running in published batches. - The plan also sets 12 headline indicators, including agent and smart-terminal adoption rising from 40% in 2025 to over 90% by 2030. ## What the plan is, and what it is not A five-year informatics plan is a direction document. It is issued by the Central Cyberspace Affairs Commission, it is binding on the agencies and local governments that must implement it, and it does not by itself create a new obligation for a company. Obligations arrive later, through departmental rules, standards or filing notices that cite the plan as their mandate. That distinction is easy to lose. Reading the plan as though agent filing were already mandatory would be wrong; reading it as irrelevant would also be wrong, because Chinese regulators rarely write a task into a five-year plan without following it with an instrument. The honest position is that the intent is now on the record and the mechanism is not. ## The filing regime that already exists China's generative AI filing system (生成式人工智能服务备案) has been operating for some time, with the cyberspace regulator publishing batches of filed services. It has also published filing information covering on-device generative AI services on phones, which is evidence that the on-device category is not waiting for the plan — it is already being processed in some form. The already-in-force rule that most resembles what the plan describes is the 《人工智能拟人化互动服务管理暂行办法》, jointly issued by the cyberspace regulator, the development and reform commission, MIIT, the ministry of public security and the market regulator. It was published on 10 April 2026 and took effect on 15 July 2026. It applies to services that simulate a natural person's personality, thinking patterns and communication style to provide ongoing emotional interaction, and explicitly excludes customer service, knowledge Q&A, work assistants, education and research where no ongoing emotional interaction is involved. Under that measure, providers must run a security assessment when launching a service, when adding relevant features, when registered users exceed 1 million, when monthly active users exceed 100,000, or when a significant security risk exists. They must also complete filing under the existing algorithm recommendation rules. Those thresholds are concrete, dated, and legally binding today — unlike the plan's one-line commitment. ## Why agents and on-device models are the hard case The existing system was built around a service that a provider operates and can be asked to modify. An agent complicates that: it calls external tools, acts rather than only answers, and its behaviour depends on what it is connected to. An on-device model complicates it differently: there may be no server-side inference to inspect, and updates arrive through app stores rather than through a service the regulator can reach. The plan gestures at both problems. Beyond filing, it calls for scenario guidance for key use cases, a maturity assessment system for agent technology, and standards for agent identity identifiers. It also proposes a security risk assessment system for applications in key sectors. None of these exist as published documents yet. ## Honest limitations I am reading the plan through the regulator's announcement, its press Q&A and the circulated full text, not through any implementing rule — because none has been published. Several numbers in circulation about the plan come from secondary summaries; I have used only the indicators confirmed in the regulator's own Q&A and official reporting. I cannot say what an agent filing will require, how long it will take, or whether on-device models will be filed by the model, the app or the device maker, because the plan does not say. The plan is also a Chinese domestic document; I have not assessed how it interacts with export or cross-border rules, which run on separate tracks. And a plan sets direction for five years — agencies sometimes quietly drop tasks that turn out to be impractical, so "written in the plan" is a strong signal, not a guarantee. ## What readers can do now - If you run an AI product in China, determine which regime already covers you. An emotional-interaction companion is already inside a measure that took effect in July 2026, with filing and assessment triggers at 1 million registered users or 100,000 monthly actives. That is not a future problem. - Keep an evidence trail now: model version, training-data provenance, evaluation results and safety test records. Every filing regime China has built asks for roughly this set, so assembling it early costs less than retrofitting it later. - Watch the cyberspace regulator's filing announcements rather than the headlines. When agents and on-device models move from plan language to procedure, it will appear first as a filing notice or a draft rule, not as a speech. ## Reading the trend The interesting move here is not the targets. It is the recognition that the unit of regulation is shifting: from "a model you serve" to "a capability that runs somewhere and acts on things". Filing is the instrument China reaches for when it wants to know who is accountable for a capability, and this plan says the category of accountable parties is about to widen. According to Li An, chief scientist at BrainNet (脑机网), the shift from filing a service to filing a behaviour is what distinguishes this round of Chinese AI governance from the generative AI rules that came before it. Ninety percent adoption by 2030 may or may not happen. The paperwork is the part that is certain. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-informatics-plan-ai-filing Free to quote with attribution and a link to the original.