In October 2025 China’s State Administration for Market Regulation and the national standards authority released a batch of AI national standards, including AI Risk Management Capability Assessment, AI Federated Learning Technical Specification, AI Deep Learning Compiler Interface, AI Computing Centre Computing Capability Assessment, and General Technical Requirements for Customer-Service Virtual Digital Humans. All took effect on publication.
What a batch release means
A single standard is a technical matter; a batch release is a systems matter. This batch spans several links in the AI value chain: risk management, data collaboration, the underlying toolchain, compute infrastructure and a specific application form.
Seen from the standards system, earlier AI standards concentrated on safety floors, such as the basic security requirements for generative AI services. This batch extends towards capability assessment and engineering specification. See our note on the generative AI security standard.
Four directions worth watching
Risk management capability assessment. The value of such standards is turning “does the company have risk management capability” from a subjective judgment into an assessable indicator, with indirect effects on regulatory inspections, procurement qualifications and findings of fault in litigation.
Federated learning technical specification. Federated learning allows joint model training without data leaving its domain, one technical path around data silos and compliance conflict. Standardisation provides engineering consensus for “usable but not visible” data, lowering the difficulty of compliance argument. See our note on data element market allocation.
Deep learning compiler interfaces. Unifying toolchain standards directly affects the cost of moving models across hardware platforms, and is key to lowering adaptation costs in domestic compute ecosystems.
Customer-service virtual digital human requirements. This is an application-level standard, pulling virtual digital humans from marketing concept back to testable indicators. Digital humans involve appearance, voice, interaction and content generation, so their compliance boundary is closely tied to labelling duties. See our note on rules for anthropomorphic interaction services.
Practical suggestions
First, put the standards list into the product compliance baseline. Most are voluntary, but they bind in practice through procurement requirements, regulatory reference and judicial assessment.
Second, distinguish floor standards from capability standards. Floor standards decide whether a product can launch; capability standards affect access to particular markets and projects. See our note on the AI application ethics and security guideline.
Third, track version iterations. AI standards update far faster than traditional industries, so a tracking mechanism beats a one-off cross-check.
A maturing standards system means AI compliance work is expanding from reading regulations to aligning with engineering specifications.
This article is based on the AI-related national standards released by China’s State Administration for Market Regulation and the national standards authority, and on public information. It is for general reference only and does not constitute legal advice.