In March 2025 the State Administration for Market Regulation disclosed progress on AI national standards. In 2024 SAMR and the relevant departments issued the guidelines for building a comprehensive standardisation system for the national AI industry; since the start of 2025 a batch of AI national standards has been published and the skeleton of the standards system is taking shape.
Three directions already published
From public information, the standards published since the start of 2025 cluster in three areas: general requirements for AI large models, evaluation metrics and methods, and service capability maturity assessment.
Service capability maturity assessment corresponds to GB/T 45907-2025, which provides a framework for grading and assessing service capability; the basic security requirements for generative AI services correspond to GB/T 45654-2025. What these standards have in common is that they convert abstract principles into testable indicators.
Progress on compute platform standards
On compute platforms, standards work is focused on deep learning compilers, high-quality datasets, and compute scheduling and coordination. That orientation answers the action plan on compute interconnection: cross-entity, cross-architecture and cross-region compute scheduling presupposes standardised interfaces and identifiers.
See the action plan for building high-quality industry datasets.
The relationship between standards and law
A national standard is not an administrative regulation and normally has no direct punitive effect. In practice, however, standards acquire binding force in three ways: they are referenced by laws and regulations and become a compliance basis; they become scoring items in tendering and government procurement; and they become a technical reference in judicial decision-making.
See our notes on AI national standards in October 2025; the update cycle of the standards system is speeding up.
How companies can keep pace
First, put standard numbers into product documentation. Citing the corresponding standard in filing materials, security assessment reports and technical descriptions helps the credibility of those materials.
Second, distinguish mandatory from recommended standards. GB and GB/T do not have the same effect; mandatory standards that touch a security floor need to be aligned first.
Third, watch the interaction between standards and filing review. See our notes on generative AI filing and registration; in review practice, judgements about technical indicators often refer to the standards in force.
The value of a standard is that it turns “we did it” into “we did it in a verifiable way”.
This article is based on the State Administration for Market Regulation’s disclosure on progress in AI national standards and public information and is for reference only; it does not constitute legal advice.