On 17 January 2024 the Ministry of Industry and Information Technology published for comment the draft National Comprehensive Standardisation System Construction Guide for the AI Industry. The targets are specific: by 2026, more than 60 per cent of generic key technology and application development projects should produce standards, and more than 50 national and industry standards should be newly developed. The system has seven parts: foundational common standards, foundational support, key technologies, intelligent products and services, enabling new industrialisation, sector applications, and security and governance.
For AI companies the question is not whether to engage with standards but where standards sit in the regulatory chain.
The seven parts order the standards
The seven parts are not a simple taxonomy. They cut the value chain into segments: foundational common standards cover terminology, reference architecture and evaluation methods; foundational support covers compute, data and tooling; key technologies correspond to algorithms and model capability; intelligent products and services cover concrete forms; enabling new industrialisation and sector applications cover deployment; security and governance stand apart. A single company may fall into several parts at once, so it needs to know its own standard coordinates.
Without that map, investment disperses. A model company joining terminal standards, or an application company chasing infrastructure standards, is unlikely to produce anything usable.
Standards, filing and security assessment
Launching a generative AI service involves filing and security assessment. A standard is not itself a licensing condition, but it supplies the technical yardstick by which regulators judge whether requirements are met. If the security measures, test methods and provenance statements in an assessment file map onto published or draft standards, communication costs fall.
Where standards are missing, regulators rely on case-by-case judgment and companies find it harder to predict. That is why participation has practical value for product planning.
Routes for participation
Participation usually runs from observer to working group member to lead drafter, with rising commitment at each step. A company can start where it has the deepest experience: model evaluation for a particular sector, data labelling quality, or security requirements for on-device inference. Technical proposals, test data and cases carry more weight than a name on a list.
There is tension between standard-setting and trade secret protection. Which details may be published and which should remain internal know-how must be decided before submission. Once a standard text is public, later claims to confidentiality have little room.
Standards as compliance evidence
In contracts, tenders and customer due diligence, standards conformity is becoming a verifiable qualification. Third-party test reports, declarations of conformity and internal test records form an evidence chain. For exporters, the same chain can answer technical questionnaires from overseas customers.
A standard is not a shield. Conformity does not eliminate liability, but it can show that reasonable care was taken when a dispute arises. The risk to avoid is treating the standard as marketing material while being unable to find the corresponding test records in the project file.
For filing and security assessment preparation, see our algorithm filing service; for security measures and graded protection, see cybersecurity compliance; terminology is collected in the glossary.
This article is based on the draft National Comprehensive Standardisation System Construction Guide for the AI Industry published for comment by the Ministry of Industry and Information Technology on 17 January 2024, and on public information. It is for reference only and does not constitute legal advice.