China AI Safety Governance Framework 3.0

Overview

人工智能安全治理框架3.0, translated as Artificial Intelligence Safety Governance Framework 3.0, was released on 14 September 2026 by the National Technical Committee on Cybersecurity Standardization. The launch took place at the opening ceremony of the 2026 National Cybersecurity Awareness Week. The official release announcement identifies the committee as publisher and states that preparation took place under the guidance of the Cyberspace Administration of China. The framework addresses new trends in artificial intelligence development and new safety governance challenges. Its objectives are to strengthen governance consensus and improve capabilities to prevent and respond to risks, enabling artificial intelligence to benefit humanity.

Version 3.0 maintains the core logic of risk classification, technical responses, and comprehensive governance. It updates risk classifications and optimizes technical responses and governance measures. The framework is guidance-oriented and non-binding, rather than legislation or a regulation establishing enforceable obligations. Its In Force status denotes publication of guidance, not commencement of binding legal duties. The announcement specifies no separate effective date, legislative adoption procedure, statutory authority for imposing duties, or enforcement mechanism. Publication therefore does not establish that organizations became subject to new legally enforceable requirements on 14 September 2026.

Definitions and Scope

Risk classification, technical responses, and comprehensive governance are the framework’s three organizing concepts. The release announcement identifies these components but does not formally define them or reproduce the classification structure. It states that classifications were updated and that technical and governance measures were optimized and adjusted. It does not name individual risk classes, assign systems to categories, identify particular controls, or prescribe institutional procedures. These concepts describe the framework’s approach without establishing specific technical requirements through the announcement alone.

The guiding principles include people-centered development, an orientation toward positive and beneficial outcomes, strengthened risk awareness, and assurance of safety and controllability. The announcement presents these as principles, not measurable compliance thresholds. It does not formally define artificial intelligence, AI provider, developer, deployer, high-risk system, or safety incident. It also identifies no exclusions, territorial application rules, or covered commercial activities. General policy terminology consequently does not establish a defined legal scope or independent legal consequences for particular actors or activities.

Governance and Institutional Framework

The National Technical Committee on Cybersecurity Standardization published version 3.0 and organized its research and preparation under the guidance of the Cyberspace Administration of China, also known as CAC. Named participants include the China Academy of Cyberspace Studies and CAC’s Data and Technical Support Center. The announcement also refers collectively to professional institutions, research institutes, and industry enterprises. Preparation therefore involved governmental guidance, technical coordination, research expertise, and industry participation. The announcement does not list every contributing organization or allocate distinct implementation responsibilities among participants.

Publication, guidance, and participation are distinct from enforcement authority. These roles do not themselves establish powers under the framework to inspect organizations, impose sanctions, approve AI systems, or adjudicate complaints. The announcement creates no new regulator, supervisory board, licensing office, or complaints mechanism. It does not specify reporting relationships for organizations using the framework or establish a compulsory oversight structure. The institutional arrangements identify the publisher, guiding authority, and contributors without assigning additional statutory responsibilities to them.

Key Focus Areas

The framework focuses on the relationship between evolving AI risks and appropriate safety governance responses. Version 3.0 reviews and updates risk classifications in light of changing circumstances and optimizes technical responses and comprehensive governance measures. These changes respond to new development trends and safety governance challenges while retaining the established relationship between classification, technical measures, and governance. The announcement does not enumerate revised risk classes, identify severity levels, or explain which measures correspond to particular risks.

The framework also addresses the direction of AI development. Its principles emphasize people-centered values and beneficial outcomes alongside risk awareness, safety, and controllability. Its objectives combine broader governance consensus with improved risk prevention and response capability. These themes concern governance oversight, risk management, and cybersecurity or model security, without establishing specific testing obligations in the announcement. The announcement does not expressly identify agentic AI, embodied AI, prohibited applications, content-labeling rules, training-data restrictions, or sector-specific requirements. Provisions of other AI instruments do not automatically become provisions of version 3.0.

Implementation Framework

The release announcement presents a structural approach rather than an operational rulebook or implementation schedule. It confirms changes to risk classification, technical responses, and comprehensive governance but does not reproduce procedures for applying those components. No mandatory sequence of assessment, approval, deployment, or reassessment is specified. Nor are duties assigned to particular categories of developers, service providers, users, or public bodies. High-level principles therefore remain distinct from detailed organizational obligations.

The official publication links to the Artificial Intelligence Safety Governance Framework 3.0 document. The announcement does not establish a required organizational implementation plan, prescribed documentation format, or certification process. Voluntary use of the framework must be distinguished from obligations arising under separately applicable binding instruments. Practical checks concerning instrument identification, publication status, institutional roles, principles, and operational details are interpretive safeguards, not compulsory actions created by the release announcement.

Monitoring and Evaluation

Version 3.0 seeks to improve capabilities to prevent and respond to AI risks and reflects attention to changing technologies and safety governance challenges. Those objectives make continuing awareness of technological developments relevant, but the announcement prescribes no monitoring program. It specifies no reporting frequency, incident notification deadline, audit interval, performance indicator, or responsible evaluation body. It also identifies no methodology for determining whether organizations have implemented the framework successfully or improved risk prevention capabilities.

The framework has developed through successive versions: version 1.0 in 2024, version 2.0 in 2025, and version 3.0 in 2026. The official version 2.0 announcement describes tracking changes in risks and updating governance measures. This history does not establish a mandatory annual review cycle for version 3.0. No next revision date, organizational reassessment deadline, compulsory benchmark, independent evaluation requirement, or public monitoring report is announced. The general objective of stronger governance and risk-response capability does not itself supply these mechanisms.

Penalties, Liability, and Appeals

The release announcement specifies no fines, criminal penalties, administrative sanctions, civil liability rules, or compensation mechanisms. It identifies neither conduct constituting a violation of the framework nor an authority responsible for investigating such violations. Version 3.0 therefore has no penalty schedule or independent enforcement regime identified in the announcement. Risk awareness, safety, and controllability are governance principles, not punishable offenses or directly enforceable legal commands established by that publication.

No appeal procedure, administrative reconsideration route, judicial review mechanism, or complaint-handling process is described. This does not determine whether separate laws impose obligations or provide remedies for a particular AI deployment. Liability, financial exposure, enforcement thresholds, and remedies under other instruments require identification of the applicable binding provisions and relevant facts. The framework’s broad principles do not substitute for those provisions or independently establish legal consequences.

Relationship to Other Instruments

Version 3.0 follows Artificial Intelligence Safety Governance Framework 1.0 and Framework 2.0, published in 2024 and 2025 respectively. The announcement connects this work to implementation of the Global AI Governance Initiative. Version 3.0 continues the established core logic while revising risk classifications and response measures. This is successive development within the same governance initiative, but the announcement supplies no formal repeal clause, transitional provision, or statement that every recommendation in earlier versions has ceased to apply.

The official version 2.0 announcement describes that earlier framework as a technical document of the National Technical Committee on Cybersecurity Standardization and discusses developments in risk classification and governance measures. Those earlier details are historical context, not automatically provisions of version 3.0. The 3.0 announcement identifies no amendments to particular statutes, incorporation into binding regulations, exemptions from other requirements, or order of legal precedence. Its publication does not establish changes to substantive obligations under separate AI, data, or cybersecurity instruments.

International Alignment

The expressly identified international policy connection is implementation of the Global AI Governance Initiative. The release links the framework’s successive versions to that initiative and emphasizes that AI should benefit humanity. People-centered development, beneficial outcomes, safety, and controllability are the stated international-facing principles. The announcement establishes no treaty obligations, mutual recognition arrangements, cross-border assessment procedures, or equivalence with another jurisdiction’s AI governance system. It identifies no international institution responsible for implementing or supervising version 3.0.

The Ministry of Foreign Affairs published the Global AI Governance Action Plan on 26 July 2025. That separate policy instrument discusses risk assessment, standards, and international cooperation on AI safety governance. The version 3.0 announcement does not state that those provisions are incorporated into the framework. The action plan is also distinct from the Global AI Governance Initiative expressly named in the release. International alignment consequently concerns a stated policy connection and objective, rather than harmonized technical requirements, internationally recognized certification, or new cross-border compliance duties.

Implementation Timeline

Compliance Checklist

The following checks support accurate interpretation and voluntary use. They are not compulsory obligations established by the framework’s release announcement.

Sources and References