China GB/T 45654-2025 Basic Security Requirements for Generative AI Services
Clause 5: model security – China GB/T 45654-2025 Basic Security Requirements for Generative AI Services

China GB/T 45654-2025 Basic Security Requirements for Generative AI Services 5.2: Model output security: ninety per cent pass rate, accuracy, reliability, refusal behaviour and content labelling

What does 5.2 require?

The provider must ensure that the pass rate of model-generated content is not below ninety per cent (the share of samples free of the 31 Appendix A risks, tested by the method in Appendix B.2.2); must take technical measures to improve the accuracy of generated content (responsiveness to the user's intent and consistency of data and statements with scientific common sense and mainstream understanding) and its reliability (sensible format and structure and useful content); must refuse to answer questions that are plainly extreme or plainly designed to induce illegal or unhealthy output, while answering all other questions; and must label generated images, video and similar content as required by national regulations and standards (the Measures for Labelling AI-Generated Synthetic Content and GB 45438-2025).

Maintained by Gerard Blokdyk

Other controls in Clause 5: model security – China GB/T 45654-2025 Basic Security Requirements for Generative AI Services

Query this from an agent

The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.