⚡ Quick Summary
Published by The Alan Turing Institute, this country profile is part of the AI governance around the world series and examines China's approach to AI regulation and standardisation from July 2017 to July 2026, drawing exclusively on primary sources.
The profile traces the policy trajectory from the 2017 New Generation Artificial Intelligence Development Plan (AIDP), with its three-step roadmap for 2020, 2025 and 2030 and six overarching goals, through the 2019 Governance Principles, the 2021 Ethical Norms, the 2023 Global AI Governance Initiative and the 'AI Plus' initiative, whose August 2025 directive sets a 70% penetration target for intelligent terminals and autonomous agents by 2027, 90% deployment by 2030 and full integration by 2035. It describes application-specific rules on algorithmic recommendations, deep synthesis, generative AI services, AI-generated content labelling and anthropomorphic interaction services, alongside Shanghai and Shenzhen local regulations.
On standardisation it sets out the SAC-coordinated, hierarchical system, the role of CESI and MIIT, SAC/TC260's Basic Safety Requirements for generative AI and AI Safety Governance Framework, more than 50 published national AI standards as of June 2026, and the seven core domains of the 2024 standard system guidelines. The conclusion characterises China's approach as incremental and adaptive, with departmental rules first and a comprehensive AI law still in progress.
🧩 What’s Covered
The profile follows a consistent framework across three main parts, from policy aims to regulation and standardisation.
- Executive summary and timeline: Summarises China's position as one of the largest AI economies and an early mover in AI regulation, and lists eight milestones from the July 2017 AIDP to the July 2026 Interim Measures for AI Anthropomorphic Interaction Services.
- High-level aims and principles: The AIDP's three-step roadmap and six goals; the 2019 Governance Principles with eight principles including harmony, fairness, privacy, safety and agile governance; the 2021 Ethical Norms; the 2023 Global AI Governance Initiative; and the 'AI Plus' deployment milestones for 2027, 2030 and 2035.
- Definitions of relevant technologies: The absence of a single AI definition, with instrument-specific definitions of 'algorithmic recommendation technology', 'deep synthesis technology' and 'generative AI technologies', plus footnote references to a draft machine learning algorithm standard and a 2018 SAC white paper definition.
- National regulatory initiatives: Requirements under the four application-specific regulations and the 2025 labelling measures, including CAC registry filing, risk assessments, news service licensing, identity verification, complaint handling, explicit and implicit content labels, and prohibitions on systems designed to induce emotional dependence or psychological control.
- Local regulatory initiatives: Shanghai's anticipatory mechanisms (sandboxes, municipal advisory committee, 'necessity, appropriateness and controllability') and Shenzhen's financial support, testing and certification route for low-risk services in the absence of standards, plus a municipal AI Ethics Committee.
- Main features of the standardisation system: SAC's position within SAMR, the national, sectoral, local, association and organisation levels, mandatory versus recommended standards, voluntary standards cited as de facto requirements, and the government's direct involvement and use of quantitative targets.
- National AI standardisation activities: CESI and MIIT, the 2020 and 2024 standard system guidelines, the seven core domains and the shift from 'safety/ethics' to 'safety/governance', TC260's Basic Safety Requirements with an annex of more than 30 risks, the AI Safety Governance Framework, and the new MIIT technical committee established in December 2024.
- International engagement and conclusion: Quantitative indicators of influence in international SDOs, and a conclusion on the agility and the legal-hierarchy limits of departmental rules.
💡 Why it matters?
For teams operating, assessing or auditing AI services with a Chinese dimension, the profile identifies which instrument applies to which application and what it obliges: registry filing for recommendation algorithms, identity verification and labelling for deep synthesis, security assessments and complaint mechanisms for generative AI, mandatory provenance labelling for synthetic content, and design constraints for companion and conversational agents. It also shows where technical requirements come from, since SAC/TC260 documents support the implementation of binding measures. The document frames itself as a foundation for comparative analysis and future work on regulatory interoperability, useful where obligations overlap with other regimes.
❓ What’s Missing
The series states that it does not comment on the efficacy of the governance models it describes, so no assessment of how rules work in practice, of enforcement, or of penalties appears. A comprehensive AI law and rules on government data-sharing are noted as still being developed, leaving the framework incomplete. Civil society's involvement in standardisation is described as not readily available. The local section deliberately covers only local regulations, setting aside policy guidelines and action plans. The profile ends at July 2026, and page 23 retains an unresolved 'Pull out quote example' placeholder.
👥 Best For
Compliance and policy analysts deciding which Chinese instruments apply to a specific AI service and what they require; standards specialists tracking SAC, TC260 and MIIT activity; product teams building synthetic media, recommendation or companion AI for the Chinese market; and comparative governance researchers working on regulatory interoperability across jurisdictions.
📄 Source Details
AI governance around the world: China, published by The Alan Turing Institute, dated July 2026 on the cover and cited as (2026). Authors Arcangelo Leone de Castris, Rhoda Jiang and Miaowei Wang, with expert review acknowledged from Jason Zhu (Concordia AI). The profile runs to 24 pages; the citation prints a DOI, https://doi.org/10.5281/zenodo.21031557. The input was a text extraction of all 24 pages, including the table of contents, executive summary, timeline, footnotes and conclusion.
