NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's AI training data now has a security rulebook — for the people who label it

GB/T 45674-2025 sets security rules for the platforms, rules, personnel and verification behind AI training-data annotation in China's generative-AI supply chain.

2026-11-14 · 880 words · NeuroAI
China's AI training data now has a security rulebook — for the people who label it

A model's worst outputs often trace back to a single mislabeled image or a poisoned annotation.

Behind every large model sits an army of human labelers that few users ever think about.

On November 1, 2025, China put that hidden layer under a formal security rulebook.

Key takeaways

  • GB/T 45674-2025, Cybersecurity Technology — Generative AI Data Annotation Security Specification (网络安全技术 生成式人工智能数据标注安全规范), is a recommended national standard released on April 25, 2025 and effective on November 1, 2025.
  • It covers four areas: security of the annotation platform or tool, security of the annotation rules, requirements for annotation personnel, and verification of the annotation work.
  • It applies to organizations that run training-data annotation, and gives data buyers and third-party assessors a reference for checking annotation security.
  • It sits beside two sibling standards on pre-training and optimization-training data security (GB/T 45652-2025) and on generative-AI service security (GB/T 45654-2025), all effective the same day.
  • Because it is a recommended, not mandatory, standard, it guides procurement and audits rather than compelling behavior on its own.

Why labeling is a security surface

Training data is where most AI safety problems are born. A subtle mislabel can teach a model the wrong correlation; a deliberately poisoned batch can open a backdoor; a careless annotation tool can leak personal information. For years the conversation about AI governance focused on the model and the output. The annotation step — the human and software pipeline that turns raw data into "ground truth" — was treated as an operations detail, not a compliance boundary.

China's cyber-security standardization body now treats it as both. GB/T 45674-2025 is part of a cluster of 2025 standards that pin down the security expectations for generative AI along the whole lifecycle, from raw data to delivered service.

What the standard actually covers

The standard organizes its requirements into four blocks.

The platform or tool block asks that the environment where labeling happens be secured — access control, data handling, and the integrity of the tooling itself, so that the act of labeling does not become a leak or tampering point.

The annotation rules block is about the instruction set: how labels are defined, who approves them, and how consistency is kept across large and often outsourced workforces. Bad or contradictory rules are a quiet source of model failure.

The personnel block sets expectations for the people doing the labeling — their vetting, training and accountability — which matters because annotation is still heavily human in industries from medical imaging to autonomous driving.

The verification block describes how annotation security should be evaluated, giving organizations a self-check and giving buyers and third-party assessors a yardstick.

Together the four blocks read less like a technical spec and more like an internal-control framework for a part of the AI supply chain that was largely unregulated on paper.

Where it sits in the AI-safety standard stack

GB/T 45674-2025 did not arrive alone. On April 25, 2025, the same batch of approvals published GB/T 45652-2025 on the security of pre-training and optimization-training data, and GB/T 45654-2025 on basic security requirements for generative-AI services. The three share a publication and effect date and a common sponsor, the National Cybersecurity Standardization Technical Committee (全国网络安全标准化技术委员会).

The logic is layered: 45652 guards the data going in, 45674 guards the human process that shapes it, and 45654 guards the service that comes out. None of the three is a mandatory national standard; all are recommended (推荐性), which means they function as the technical backbone for procurement, self-assessment and third-party evaluation rather than as standalone legal commands.

What it can and cannot do

The standard is strongest as a procurement and audit reference. A company buying annotated data, or a regulator commissioning a security review, now has a named document to point to. It is weaker as enforcement: being recommended, it does not by itself create a penalty, and it does not specify how small annotation shops — where much of the real work happens — should prove compliance.

Its arrival also signals a shift in emphasis. Instead of only asking "is the model safe," Chinese standard-setters are increasingly asking "was the data behind it handled safely," and they are writing that question into numbered clauses.

Honest limitations

This article covers one recommended national standard and its two closest siblings; it does not track sector-specific annotation rules, provincial implementation guides, or how certification bodies actually test against it. Being a recommended standard, its real-world bite depends on whether buyers and regulators adopt it in contracts and audits, which we did not survey. We also did not inspect the full text clause-by-clause, so the four-block summary reflects the published scope description rather than a line-item reading. Figures such as publication and effect dates come from the national standards platform and the standardization committee's own bulletins.

What readers can do now

  1. If you buy or commission annotated training data in China, add GB/T 45674-2025 as an explicit acceptance criterion and ask vendors how they meet its personnel and verification blocks.
  2. If you run an annotation operation, map your platform access controls, rule-approval workflow and worker vetting to the four blocks before a client audit lands.
  3. If you evaluate AI vendors, treat annotation-security documentation as a due-diligence item on par with model-card and privacy paperwork, not as a back-office formality.

Related coverage

Cite this article

NeuroAI (脑机网), "China's AI training data now has a security rulebook — for the people who label it," 14 November 2026, https://neuroai.site/a/na-policy-ai-data-annotation-security-stand

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories