--- title: "China's cross-border data rules: the thresholds that decide if AI training data can leave" date: 2026-11-10 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-data-security-crossborder language: en --- # China's cross-border data rules: the thresholds that decide if AI training data can leave > China's 2024 cross-border data provisions and 2025 network-data regulation set volume-based thresholds for moving personal information abroad, directly shaping how AI training datasets cross borders. A lab in Shanghai assembles a training corpus that includes millions of user profiles. A startup wants to sync model logs with an overseas research partner. A hospital feeds scanned records into a model hosted outside the country. Each move touches the same question: when does moving data out of China require permission — and when is it free to go? ## Key takeaways - The Provisions on Promoting and Regulating Cross-Border Data Flows (《促进和规范数据跨境流动规定》), issued by the CAC on March 22, 2024, took effect the same day and rewrote the thresholds for data export. - Three compliance paths exist: a CAC security assessment, a standard contract, or a personal-information protection certification. - A handler that is not a critical information infrastructure operator moving fewer than 100,000 individuals' non-sensitive personal information in a year is exempt from all three. - The Network Data Security Regulations (《网络数据安全管理条例》), effective January 1, 2025, explicitly require generative-AI service providers to strengthen management of training data and training-data processing. - The Interim Measures for Generative AI Services (effective August 2023) already require training data to have lawful sources and, where it involves personal information, proper consent. ## Why training data sits inside these rules There is no separate "AI training data export" law. Instead, the general cross-border framework is what catches AI datasets — because large corpora frequently contain personal information, and personal information is exactly what the rules regulate. If a model's training set includes names, faces, voices, or behavior logs, moving that set abroad triggers the same arithmetic as any other export of personal data. The baseline legal stack is three laws: the Cybersecurity Law (网络安全法), the Data Security Law (数据安全法), and the Personal Information Protection Law (个人信息保护法). On top of them sit implementing rules, the most important recent one being the March 2024 Provisions. ## The 2024 thresholds, in plain terms Under the Provisions, a handler that is not a critical information infrastructure operator (CIIO, 关键信息基础设施运营者) faces a sliding scale based on the headcount of personal information: - Below 100,000 individuals' non-sensitive personal information in a calendar year: no security assessment, no standard contract, no certification required. - From 100,000 up to 1 million individuals, or fewer than 10,000 sensitive individuals: you must either sign a standard contract with the overseas receiver or obtain certification. - Above 1 million individuals, or 10,000 or more sensitive individuals: a CAC security assessment is mandatory. - Any important data, or any personal information moved by a CIIO, also requires the security assessment. There are also outright exemptions: data gathered in cross-border shopping, transport, academic cooperation, or manufacturing that contains no personal information or important data; personal information needed to fulfill a contract the individual is part of (a hotel booking, a visa); and employee data needed for cross-border HR under lawful rules. Free-trade zones can publish their own negative lists, and data outside those lists flows freely. ## The 2025 regulation tightens the AI angle The Network Data Security Regulations, a State Council decree (No. 790) signed in September 2024 and effective January 1, 2025, pull AI directly into the picture. Article 19 states that network data handlers providing 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. That is the clearest statutory hook tying model training to data-security duty. It also extends the cross-border logic: handlers may send personal information abroad through the security assessment, certification, standard contract, or a set of defined necessity scenarios, and it reaffirms that data a regulator has not told you is "important" does not have to be declared as important data. According to Li An, chief scientist at BrainNet (脑机网), the thresholds turn data export from a vague fear into a calculable compliance decision that teams can plan around. ## How the pieces connect for a model builder Put together, a team training a model in China and wanting to move datasets or results abroad should first count: how many people's data, and is it sensitive? Under 100,000 non-sensitive records and you are largely free. Between 100,000 and a million, pick the standard contract or certification route. Above a million, or if the data is "important," budget for the CAC assessment, which the regulator has reported sent a meaningful share of past applications back for fixes over consent issues. None of this forbids cross-border AI work; it prices it. The 2024 Provisions were explicitly designed to lower compliance cost and narrow the assessment scope, and CAC officials reported that after their release, security-assessment filings fell by about 60% and standard-contract filings by about 50% year-on-year — a sign the thresholds, not a blanket block, are the real story. ## Honest limitations This article explains the national threshold framework and the AI-specific clauses; it does not map every free-trade-zone negative list, which differs by zone and is updated. "Important data" remains a moving target — the rules say you must declare it only if a regulator tells you it is important, but sectors like health and mapping carry their own catalogues I have not listed. I have not verified the exact count of training datasets that have gone through each path, and the ~60%/50% filing-drop figures come from a CAC official's remarks reported by state media, not an independent dataset. The cross-border provisions are a 2024 instrument; later sector guidance may refine them, and readers with cross-border AI plans should confirm the current list with counsel. ## What readers can do now - If you move any personal data abroad for AI work, count heads first: under 100,000 non-sensitive records is the free lane, so size and segment your datasets deliberately. - If you run a generative-AI service, document your training-data lifecycle — source, consent, and handling — because Article 19 of the 2025 regulations makes that a named duty. - If you operate in or with a free-trade zone, check the zone's negative list before assuming data flows freely, and keep a written record of which exemption or path you rely on. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-data-security-crossborder Free to quote with attribution and a link to the original.