China has finalized its national safety standard for generative artificial intelligence services, establishing comprehensive and binding requirements for AI developers and service providers operating in the country. The standard, designated GB/T 45654–2025, was translated and analyzed by Georgetown University’s Center for Security and Emerging Technology (CSET), providing an authoritative English-language window into Beijing’s evolving approach to AI governance.
The finalized standard establishes enforceable baselines covering critical aspects of AI development and deployment, including training data sourcing and vetting, model safety testing protocols, and content moderation mechanisms. This regulatory framework is designed to ensure that all generative AI services operating within China adhere to strict safety requirements and align with the government’s broader policy objectives for the technology.
(Related: China’s AI Agent Regulation Debate Enters a New Era as Hangzhou Tests Autonomous Governance)
What the Standard Requires
Under GB/T 45654–2025, AI service providers must implement rigorous vetting processes for their training datasets to prevent the inclusion of illegal, harmful, or politically sensitive content. The standard also mandates robust model safety testing to mitigate risks such as algorithmic bias, the generation of inappropriate outputs, and the potential for AI systems to be manipulated into producing harmful content.
Content moderation is a central pillar of the regulation. Providers are required to deploy effective filtering systems to monitor and manage user interactions in real time, ensuring that AI-generated outputs comply with Chinese law and regulatory expectations. The standard also addresses transparency requirements, including obligations to disclose when content has been generated by AI, a provision that aligns with China’s earlier interim rules on generative AI services, which came into force in August 2023.
The finalization of GB/T 45654–2025 represents a significant step in the maturation of China’s AI regulatory architecture. Unlike the interim rules, which were relatively high-level, the new standard provides detailed, technical specifications that give companies clearer guidance on what compliance actually requires in practice.
Implications for the AI Industry
For AI developers operating in China, compliance with GB/T 45654–2025 will require substantial investment in data curation infrastructure, safety testing pipelines, and content moderation systems. Companies that have already built robust compliance frameworks under the 2023 interim rules will be better positioned to meet the new requirements, but the additional specificity of the standard means that many providers will need to audit and upgrade their existing processes.
The standard also has implications for international companies seeking to operate in the Chinese market. Foreign AI service providers must comply with the same requirements as domestic companies, creating a level playing field in terms of regulatory burden while also raising the barrier to entry for companies that lack established compliance infrastructure in China.
(Related: China Releases National AI Metrology Roadmap to Shift Focus from Scale to Quality)
From a geopolitical perspective, the finalization of GB/T 45654–2025 signals China’s intent to establish its own comprehensive AI governance framework independent of — and potentially in competition with — emerging international standards. As the European Union’s AI Act enters its implementation phase and the U.S. continues to develop its own regulatory approach, China’s national standard adds another distinct regulatory regime to the global AI governance landscape, with significant implications for multinational companies navigating compliance across jurisdictions.
China has also been active in international AI standards bodies, including the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), where it has sought to shape global norms aligned with its domestic regulatory approach. The finalization of GB/T 45654–2025 strengthens China’s position in these forums by demonstrating that it has a mature, detailed national standard to offer as a reference point. For countries in the Global South developing their own AI governance frameworks, China’s standard — backed by the world’s largest AI deployment base — may prove an attractive model, particularly for nations skeptical of Western-led regulatory frameworks.
For AI practitioners and policy observers, the CSET translation of GB/T 45654–2025 is a valuable resource that makes China’s regulatory thinking accessible to an English-speaking audience. Understanding the specific technical requirements embedded in the standard, rather than relying on high-level summaries, is essential for companies that need to comply, for policymakers designing competing frameworks, and for researchers studying the divergence of AI governance approaches across major powers. The standard’s emphasis on training data quality and content moderation reflects distinctly Chinese regulatory priorities, but its broader framework for model safety testing has parallels with approaches being developed in the EU and UK, suggesting that some degree of regulatory convergence may be possible even amid broader geopolitical competition.