--- title: "China's Network Data Security Regulations extend rules to AI training data" date: 2026-10-27 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-network-data-security-regulations language: en --- # China's Network Data Security Regulations extend rules to AI training data > State Council Order No. 790, in force since Jan 1, 2025, governs network data handling and explicitly pulls generative AI training data under security duties. A model's quality is only as trustworthy as the data it trained on — and a regulation that took effect on January 1, 2025 now says that trust is a compliance question, not just an engineering one. The *Regulations on Network Data Security Management* (网络数据安全管理条例), issued as State Council Order No. 790 (国务院令第790号), is the first dedicated State Council-level rulebook tying day-to-day network data handling to the older Cybersecurity Law, Data Security Law, and Personal Information Protection Law. Signed by Premier Li Qiang (李强) on September 24, 2024 and adopted at the 40th State Council executive meeting on August 30, 2024, the text runs nine chapters and 64 articles. Its reach is broad, but one clause matters directly to the AI sector. ## Key takeaways - The regulations (条例) have been in force since January 1, 2025, as a State Council administrative regulation — a higher legal rank than agency measures. - Article 19 explicitly requires providers of generative AI services (生成式人工智能服务) to strengthen security management of training data and training-data processing activities. - Processors of personal information (个人信息) on more than 10 million individuals take on "important data" obligations, including annual risk assessments. - The text builds the legal bridge for cross-border personal-information transfers through assessment, certification, or standard contracts. ## A regulation, not a guideline The instrument's legal weight is the first thing to understand. Unlike an agency notice, a State Council regulation sits directly under the laws it implements and binds provinces, ministries, and courts. That elevation matters for AI companies: obligations that previously lived in softer guidance now carry the force of administrative regulation. The regulation does three structural things. It sets general network-data security duties — classification and graded protection, breach reporting, and a ban on illegal collection or trading of data. It details personal-information protection, including clearer rules on consent, sensitive-data handling, and the right to port one's data to another provider. And it builds the cross-border transfer pathway that firms with overseas operations depend on. ## Where AI enters Article 19 is the clause AI labs should bookmark. It states that network-data processors who provide generative AI services shall strengthen the security management of training data and training-data processing activities, and take effective measures to prevent and handle network-data security risks. The wording is brief but deliberate: it places the training corpus — scraping, licensing, cleansing, and augmentation — inside the regulated perimeter rather than treating models as a separate, ungoverned layer. This closes a gap. Earlier rules governed how a deployed chatbot answered; they said little about the data pipeline behind it. The regulation reframes that pipeline as network data processing subject to the same duty of care as any other handler. ## The 10-million-person line Another provision reshapes compliance planning. Under the regulation, a processor of personal information on more than 10 million individuals (1000万人以上) must also meet the obligations the text imposes on handlers of important data — appointing a network-data security officer and a dedicated management body, conducting an annual risk assessment, and reporting it to provincial-level authorities. For large model developers and consumer apps, that threshold is not distant. A single popular AI assistant can sweep past 10 million users quickly, which means training-data governance, retention, and transfer controls move from "when we scale" to "now." The annual assessment requirement also creates a recurring audit rhythm rather than a one-time clearance. ## Cross-border without guesswork The regulation codifies how personal information leaves the country. A handler may provide personal information abroad if it passes a national data-export security assessment, obtains personal-information protection certification from a professional body, or uses the standard contract mechanism set by the cyberspace authority. It also preserves carve-outs for contract performance, cross-border human-resource management, legal obligations, and emergencies protecting life, health, or property. The practical effect is predictability. A multinational running an AI product in China can plan its data flows against a published menu of routes instead of ad-hoc negotiations — provided it documents purpose, scope, and recipients as the regulation demands. ## Honest limitations The regulation is a framework, not a manual. Article 19's "strengthen security management" is a mandate without a checklist; the real operating detail arrives through implementing rules and standards that agencies issue afterward. "Important data" itself is defined through catalogs compiled by sectors and regions, so a given dataset's status may stay ambiguous until a regulator labels it. And while the text empowers penalties under existing laws, it does not invent a new fine schedule, so enforcement intensity will vary by locality and by how seriously a case touches security or scale. ## What readers can do now - If you operate a generative AI service, map your training-data lifecycle — sourcing, licensing, cleansing, and retention — against security-duty expectations before an annual assessment cycle catches you unprepared. - If your product holds personal information on more than 10 million people, stand up the security-officer and management-body structure the regulation requires and schedule the first yearly risk report. - If you move personal data across borders, pick a documented route — assessment, certification, or standard contract — and keep the purpose and recipient records the text calls for. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-network-data-security-regulations Free to quote with attribution and a link to the original.