---
title: Do China's AI content labelling rules apply to AI agents?
url: https://opencontroller.lyzr.com/blog/china-ai-content-labelling-rules-and-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Do China's AI content labelling rules apply to AI agents?

**Yes, when an agent is part of a generative AI service offered to the public in China. Since 1 September 2025, China's Measures for Labeling AI-Generated Synthetic Content require providers to add explicit (visible) labels and implicit (metadata) labels to AI-generated text, images, audio, video and virtual scenes. An agent that writes emails, posts, documents or media for people in China is generating that kind of content.**

Global companies running customer-facing agents in China ask whether labelling is a chatbot problem or an agent problem. It is both, because the rules attach to the content and the service, not to the interface.

## What do the labelling rules require?

The Measures were issued in March 2025 by the Cyberspace Administration of China (CAC), with the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. A mandatory national standard, GB 45438-2025 (Cybersecurity technology — Labeling method for content generated by artificial intelligence), took effect on the same date and sets the technical format.

| Label type | What it is | Agent example |
|---|---|---|
| Explicit label | Visible text, sound, graphic or prompt telling users content is AI-generated | A note at the start or end of an agent-drafted message or document |
| Implicit label | Information embedded in file metadata, such as content attributes, provider name or code, and a content ID | Metadata on an image or PDF the agent generates and lets users download |

Further obligations, as described in published legal summaries:

1. **Downloads and exports** of AI-generated files must carry the required explicit labels.
2. **App distribution platforms** must check whether apps offer generative AI services and review labelling materials.
3. **Users** who publish AI-generated content must declare it and use platform labelling functions.
4. **Unlabelled output on request.** Where a user asks for content without explicit labels, the provider may supply it but must set out the user's labelling duties in the user agreement and keep related logs, reported as at least six months.

## Which AI agents are in scope?

The Measures build on China's existing rules for algorithm recommendation, deep synthesis and generative AI services. In practice:

- **In scope:** agents that generate content for the public in China, such as customer-service agents, marketing content agents, and assistants embedded in consumer apps.
- **Likely in scope:** agents that produce files users can download or forward, because labels must survive export.
- **Less clear:** internal enterprise agents with no public-facing output. China's generative AI measures carve out internal use not offered to the public, but get local legal advice for your case.

Labelling is a content rule. Separate rules cover algorithm filing and security assessments for public-facing services.

## How do you implement labelling for agents?

Labelling is easiest when it sits in the request path, not in each agent's code.

1. **Classify outputs.** For each agent, list the content types it produces and where they go (screen, email, file, social post).
2. **Decide the label.** Explicit label wording and placement per content type, following GB 45438-2025.
3. **Add metadata.** Implicit labels for generated files: content attributes, provider identifier, content ID.
4. **Apply centrally.** Enforce labelling as a policy at the gateway or output layer so every agent gets it.
5. **Log decisions.** Record which outputs were labelled, and any user-requested exceptions, with retention that meets the six-month expectation. See [how long to retain AI agent logs](/blog/how-long-to-retain-ai-agent-logs).
6. **Test.** Add labelling checks to release evaluations so a prompt change does not drop the label.

## How does this compare with EU disclosure rules?

The EU AI Act's Article 50 also requires disclosure when people interact with AI and machine-readable marking of synthetic content, with its own timeline. China's rules are already in force and more prescriptive about format. A single output policy that supports both visible notices and metadata marking covers the common core; see [do AI agents need to disclose they are AI](/blog/do-ai-agents-need-to-disclose-they-are-ai-eu-ai-act). Our pillar on [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how to manage regional variants.

## How Opencontroller does this

Opencontroller puts control in the request path: policies can allow, route, redact or block traffic through the LLM Gateway and Agent Gateway. Policy as code is extensible, so teams can write output policies per region, and every decision is tied to the agent and version in the governed registry. Logs of agent actions support the record-keeping these rules expect.

## Frequently asked questions

**When did China's AI labelling rules take effect?**
1 September 2025, together with the mandatory national standard GB 45438-2025.

**Does an AI agent's text output need a label?**
If the agent is part of a generative AI service offered to the public in China, AI-generated text is in scope and needs an explicit label in a visible position, such as the start or end.

**Do internal enterprise agents need labels?**
It is less clear. China's generative AI rules focus on services offered to the public. Get local legal advice for internal-only agents.

To manage regional output policies across agents, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

China's Measures for Labeling AI-Generated Synthetic Content, issued in March 2025 by the CAC with MIIT, MPS and NRTA, took effect on 1 September 2025 alongside mandatory national standard GB 45438-2025. Providers must add explicit (visible) labels and implicit (metadata) labels to AI-generated text, images, audio, video and virtual scenes, keep labels on downloads and exports, and, when supplying unlabelled content on user request, specify user labelling duties in the user agreement and retain logs (reported as at least six months). Public-facing AI agents in China that generate content are in scope; internal-only agents are less clear.

## Related

- [Do AI agents have to disclose they are AI under the EU AI Act?](https://opencontroller.lyzr.com/blog/do-ai-agents-need-to-disclose-they-are-ai-eu-ai-act.md)
- [How long should you keep AI agent logs and traces?](https://opencontroller.lyzr.com/blog/how-long-to-retain-ai-agent-logs.md)
- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
