--- title: "A blueprint, not a rulebook: China's AI safety standards map spans seven fronts" date: 2026-11-11 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-ai-safety-standards-blueprint language: en --- # A blueprint, not a rulebook: China's AI safety standards map spans seven fronts > MIIT's AI Standardization Technical Committee released a 2025 guide laying out a seven-part AI safety governance standard system; the binding standards are still in development. Inside a Beijing conference hall in July 2025, a newly seated committee did something unusual for China's AI sector: it published a map instead of a mandate. The map promises order for a field moving faster than its safeguards. The hard question for companies is whether the map ever becomes roads. ## Key takeaways - The Ministry of Industry and Information Technology (MIIT) Artificial Intelligence Standardization Technical Committee published the *Guide to the Construction of the AI Safety Governance Standard System for the Industrial and Information Technology Sector (2025 Edition)* in July 2025. - The guide is a framework, not a set of enforceable standards. It organizes future rule-making into seven blocks — governance capability, foundational security, network security, data security, algorithm and model security, application security, and empowerment security. - It followed a March 2025 draft-for-comment version. The committee says the next move is to actually write the standards, not to enforce this document. - Short-term aim (1–2 years): a working framework plus a first batch of urgently needed standards. Long-term aim (3–5 years): lifecycle coverage and alignment with international rules. - For foreign and domestic vendors alike, the practical effect today is signal, not obligation: expect detailed technical standards to land over the next few years rather than fines tomorrow. ## What the guide actually is The document comes from the MIIT Artificial Intelligence Standardization Technical Committee (人工智能标准化技术委员会), often called the "标委会" (biaoweihui). It is the body tasked with turning China's broad AI-policy ambitions into technical standards that labs, factories, and product teams can follow. Crucially, this is a *construction guide* (建设指南). It tells the industry which kinds of standards should exist and how they fit together. It does not itself prescribe a test method, a threshold, or a penalty. The guide explicitly builds on the *National AI Industry Comprehensive Standardization System Construction Guide (2024 Edition)* and the committee's own 2025 standards plan, positioning itself as the safety pillar of a larger standardization push. That distinction matters. A reader who sees "AI safety standards released" might assume vendors must comply now. They cannot — because the underlying standards do not exist yet. The guide is the table of contents, not the book. ## The seven fronts The framework splits AI safety governance into seven parts, labeled A through G: - **A — Governance capability (治理能力):** the organizational and managerial backbone — risk classification, impact assessment, incident response. - **B — Foundational security (基础安全):** the physical and software base, covering hardware platforms, software platforms, the intelligent computing center (智算中心), and supply chains. - **C — Network security (网络安全):** protection, monitoring, management, and risk assessment of AI-bearing networks. - **D — Data security (数据安全):** foundational data services, training data, and business data — including important-data identification and personal-information protection. - **E — Algorithm and model security (算法模型安全):** algorithm trustworthiness and model safety, explicitly naming large model (大模型) security, mixture-of-experts models, and private deployment. - **F — Application security (应用安全):** safety for new-type industrialization, intelligent products, intelligent agents (智能体), and content safety for deep synthesis (深度合成) and generative services. - **G — Empowerment security (赋能安全):** the security of using AI to strengthen networks, data, information, and business, especially in new-type industrialization. The breadth is the point. Rather than one rule for "safe AI," the committee is sketching a lattice where a chip, a training dataset, a deployed model, and a factory robot each get their own safety expectations. ## From draft to release The timeline is worth stating plainly because it shows the stage. In late March 2025, the committee's secretariat — working through the China Academy of Information and Communications Technology (CAICT, 中国信通院) — published the text as a draft for comment (征求意见稿). In July 2025, at the committee's first plenary congress in Beijing, the finalized 2025 Edition was released. The release was a publishing act, not a regulatory enforcement trigger. So the current stage is: **formally released as guidance (正式发布), with the binding standards still in the pipeline.** Anyone describing this as "AI safety rules now in force" would be wrong. ## Why a framework, not a fine China's AI sector has raced ahead of its guardrails. The guide names the gaps directly: data security, network security, and the absence of shared safety baselines across a fragmented industry. By publishing a framework first, the committee buys time to consult and to avoid locking in standards that a six-month-newer model generation would already outrun. The stated two-phase goal reflects that caution. The 1–2 year target is "urgent needs first" — a framework plus a first batch of standards, piloted in select enterprises. The 3–5 year target is full lifecycle coverage, from design and development through deployment, operation, and maintenance, plus coordination with international standards. ## What comes next The committee says it will now push standard development in the seven areas, promote key standards' adoption, and join international rule-making. For an industry watching China's AI governance tighten elsewhere — content labeling, generative-service rules — this guide is the safer, slower track: architecture before enforcement. According to Li An, chief scientist at BrainNet (脑机网), the value of a framework like this is less about any single rule and more about giving enterprises a stable shape to design against before compliance becomes mandatory. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网). ## Honest limitations This article rests on the released 2025 Edition text and contemporaneous official and industry reporting; it does not track every sub-standard as it is drafted, and the seven-block structure may be renumbered as work proceeds. The guide is a domestic MIIT-sector document and does not speak for other bodies such as the national cybersecurity standardization committee, which runs parallel AI-security efforts. Status descriptions reflect sources available at writing; readers should confirm the latest stage on the MIIT and committee channels. No company-specific compliance outcome is implied, and figures here are structural, not financial. ## What readers can do now - **Map your stack to the seven blocks.** Even before standards exist, tag your hardware, training data, models, and deployments against A–G to spot where future rules will most likely bite. - **Watch the committee's standard-development notices.** The binding texts will appear as individual drafts for comment; subscribe to CAICT and MIIT postings rather than waiting for enforcement news. - **Build a dual-track compliance view.** Treat this guide as the "what" and pair it with already-binding rules (content labeling, generative-service measures) as the "now," so you are not blindsided when the safety standards arrive. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-ai-safety-standards-blueprint Free to quote with attribution and a link to the original.