On 5 June 2024 the Ministry of Industry and Information Technology, the Cyberspace Administration of China, the National Development and Reform Commission and the Standardisation Administration issued the Guide to Building a Comprehensive National Standards System for the AI Industry (2024 Edition). In the same year, TC260-003, Basic Security Requirements for Generative Artificial Intelligence Services, set security requirements for corpora, models and outputs. The two documents are often cited together, but they do different work: the guide supplies coordinates, while TC260-003 supplies a checkable security baseline. What companies need is to turn both into evidence that can be submitted and verified.
The standards system provides coordinates
The guide divides the standards system into seven parts covering foundational common standards, foundational support, key technologies, intelligent products and services, enabling new industrialisation, sector applications, and security and governance. For a company, the use is not clause-by-clause comparison but positioning: where the product sits in the industrial chain, which family of standards applies, and which standards will be cited in filing or assessment. Without clear coordinates, two kinds of waste follow: joining standards work that is irrelevant, or missing the security standards that directly concern the product.
The three segments of TC260-003
The requirements can be read in three segments: corpus, model and output. The corpus segment addresses provenance, authorisation, quality and filtering; the core question is whether training data can be explained and used. The model segment addresses security measures during training and fine-tuning; the core question is whether capabilities are appropriately constrained. The output segment addresses content safety, labelling and handling mechanisms; the core question is whether problems in generated results can be detected and corrected. The three form a chain, and a missing segment undermines the whole explanation.
Test reports must answer three questions
Test reports in filing and assessment materials often state conclusions without process. A usable report answers at least three questions: what was tested, including model version, data scope and test scenarios; how it was tested, including method, sample, environment and pass criteria; and what the result means, including uncovered areas and residual risk. A report without a version number and method has almost no probative value on review. Testing need not be voluminous, but it must correspond to specific standard clauses and product functions.
Cross-corroboration between filings
Algorithm filing focuses on algorithm mechanism and function descriptions, while generative AI service filing and security assessment focus on service form and security measures. The factual basis of the two sets of materials should be consistent: the same function, the same service recipients and the same content flows. Standards matter here because they provide shared terminology and test methods, allowing descriptions across different forms to correspond. Copying standard clauses into materials without attaching test records changes the wording but adds no evidence.
Three properties of evidence
For standards adoption to become evidence, three properties are needed: traceability, so that it is clear which version of the standard and which test record are relied on; reproducibility, so that a third party can repeat the method and environment; and correspondence, so that each control measure points to a specific clause and a specific product function. A company can maintain an evidence register organised by standard clause, control measure, test record and responsible person, updated with product versions. For overseas customer questionnaires, this register is usually more persuasive than a declaration of conformity.
For standards review and filing alignment, see algorithm filing; for security measures and testing arrangements, see cybersecurity compliance; definitions are collected in the glossary.
This article is based on the Guide to Building a Comprehensive National Standards System for the AI Industry (2024 Edition), issued by four departments on 5 June 2024, and TC260-003, Basic Security Requirements for Generative Artificial Intelligence Services, together with public information. It is for reference only and does not constitute legal advice.