EU AI Act vs China’s AI Content Labeling Rules: Two Approaches to Deepfake Transparency

Deepfakes have become a governance problem for every company that generates, edits or distributes content with AI. The world's two largest regulatory blocs have now both answered it, and they've answered it very differently.

China moved first. Since 1 September 2025, its Measures for Labeling AI-Generated Synthetic Content have required almost all AI-generated content distributed to the public to carry both a visible label and an embedded machine-readable label, backed by a mandatory national standard and an active enforcement campaign. The EU followed on 2 August 2026, when Article 50 of the AI Act began to apply. It requires machine-readable marking of AI outputs and visible disclosure of deepfakes, supported by a voluntary Code of Practice and Commission guidelines.

If you serve users in both markets, you can't simply copy one compliance setup into the other. This guide compares the two regimes side by side: who they cover, what they require, how they're enforced, and how to build one labelling pipeline that satisfies both.

The short answer

China treats AI labelling as an infrastructure rule for the whole information ecosystem. Every link in the chain, from generator to platform to app store to user, has a defined duty, and the technical format is fixed by a mandatory national standard.

The EU treats AI labelling as a targeted transparency rule. Every generative AI output must be machine-detectable, but visible labels are required mainly where deception is a real risk: deepfakes and AI-generated text published to inform the public on matters of public interest. The technical method is left open, guided by a voluntary code rather than a mandated standard.

China's approach: dual labelling across the whole chain

The Measures were jointly issued on 14 March 2025 by four agencies: the CAC, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. They build on two earlier rules: the Deep Synthesis Provisions (effective January 2023) and the Interim Measures for Generative AI Services (August 2023). The companion standard, GB 45438-2025 Cybersecurity Technology — Labeling Method for Content Generated by Artificial Intelligence, sets the technical format and is mandatory.

Two types of label

• Explicit labels: visible text, sound or graphics that users can clearly perceive, for example an "AI-generated" notice on an image, a spoken disclaimer in audio, or a persistent marker in video. When content is downloaded or exported, the label must remain in the file.
• Implicit labels: metadata embedded in the file that users don't easily notice, including the content's attributes, the service provider's name or code, and a content ID. Digital watermarks are encouraged as an additional layer.

Duties at every link in the chain

One notable flexibility: a user can ask a provider for content without explicit labels. The provider may agree if the user agreement sets out the user's labelling obligations, and it must keep the related logs for at least six months. Implicit labels still apply.

Enforcement is active

China has moved quickly from rules to enforcement. The CAC published its first batch of labelling enforcement cases in November 2025, followed by further rounds in 2026. In April 2026 it launched the special campaign "Qinglang: Rectifying AI Application Irregularities", which explicitly targeted inadequate labelling of AI-generated content alongside model registration failures and data poisoning.

Figures as reported by the CAC via Xinhua on 6 July 2026 and 2 September 2026. The campaign covers several AI misuse categories, not only labelling.

The EU's approach: Article 50 transparency

Article 50 sets four transparency duties. Two of them deal directly with AI-generated content:

• Article 50(2), providers: providers of AI systems that generate synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The solutions must be effective, interoperable, robust and reliable as far as technically feasible.
• Article 50(4), deployers: deployers who generate or manipulate image, audio or video content that is a deepfake must disclose that it is artificial. Deployers who publish AI-generated or manipulated text to inform the public on matters of public interest must also disclose it.

The other two cover telling people they are interacting with an AI system, such as a chatbot (Article 50(1)), and informing people exposed to emotion recognition or biometric categorisation (Article 50(3)).

What counts as a deepfake

The AI Act defines a deepfake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. The Commission's final guidelines, published on 20 July 2026, clarify two points that matter in practice:

• Intent doesn't matter. Whether content "would falsely appear authentic" is judged by how it looks, not by whether the deployer meant to deceive.
• Obvious fantasy is out. Content that resembles nothing real, such as a talking animal or a dragon, generally isn't a deepfake.

Exceptions China doesn't have

For deepfakes that are part of an evidently artistic, creative, satirical or fictional work, disclosure is reduced, not removed: it must be done in a way that doesn't spoil the display or enjoyment of the work. For public-interest text, no label is needed where the text has undergone genuine human review or editorial control and a person or organisation holds editorial responsibility. The guidelines read that exception narrowly. A spell-check or a quick sign-off is not enough.

The Code of Practice and EU icons

On 10 June 2026, the Commission published the final Code of Practice on marking and labelling AI-generated content. It has two sections: marking and detection for providers, and labelling for deployers. It is voluntary, but it gives signatories a practical route to show compliance. The Code accepts that no single technique is enough on its own, so in practice a multi-layered approach (metadata, watermarking and other provenance signals) is the working standard. It also introduced EU icons that deployers can use for fully AI-generated and AI-modified content.

If you are mapping these transparency duties for your own products or campaigns, I post practical breakdowns of the EU AI Act, China's AI rules and ISO 42001 regularly on LinkedIn. Feel free to connect with me there and tell me which labelling question your team is stuck on.

EU AI Act vs China's labelling rules: side by side

How the two timelines unfolded

• 10 Jan 2023

  China's Deep Synthesis Provisions take effect, requiring prominent labels on deep synthesis content.

• 15 Aug 2023

  China's Interim Measures for Generative AI Services take effect.

• 1 Aug 2024

  EU AI Act enters into force.

• 28 Feb 2025

  China publishes the mandatory standard GB 45438-2025.

• 14 Mar 2025

  China releases the final Labeling Measures.

• 1 Sep 2025

  China's Labeling Measures and GB 45438-2025 take effect.

• Nov 2025

  CAC publishes its first batch of labelling enforcement cases.

• Apr 2026

  CAC launches the Qinglang campaign on AI application irregularities.

• 10 Jun 2026

  EU publishes the final Code of Practice on marking and labelling AI-generated content.

• 20 Jul 2026

  Commission publishes final guidelines on Article 50.

• 2 Aug 2026

  EU Article 50 transparency obligations apply.

• 2 Sep 2026

  Phase 2 of China's Qinglang AI campaign is announced, focusing on deepfakes, impersonation and AI falsehoods.

• 2 Dec 2026

  Deadline for EU Article 50(2) marking by generative AI systems placed on the market before 2 August 2026.

Where the two regimes agree

Despite their different philosophies, the two regimes overlap more than headlines suggest:

• Both require machine-readable signals embedded in AI outputs, not only visible notices.
• Both put the first duty on the generator, the provider of the AI system or service.
• Both expect labels to survive download, export and redistribution as far as technically possible.
• Both treat deepfakes of real people and events as the highest-risk category for deception.
• Both encourage layered techniques: metadata plus watermarking, rather than a single method.

That overlap is good news for global providers. A well-designed provenance pipeline, for example metadata with content credentials plus a robust watermark, can carry most of the load in both markets. The differences sit mainly in visible labelling scope, format specifications and platform-side duties.

How to build one labelling pipeline for both markets

1. Map your roles in each regime

   In the EU, decide whether you act as a provider (Article 50(2)), a deployer (Article 50(4)), or both. In China, decide whether you are a generation service provider, a content distribution platform, an app distributor, or a combination.

2. Embed machine-readable marks at generation

   Add metadata and a robust watermark to every output by default. Include the fields GB 45438-2025 requires for China, and use interoperable provenance formats for the EU.

3. Design visible labels by market

   For China, apply explicit labels to all public-facing AI content, using the placement and format rules in GB 45438-2025. For the EU, apply clear labels or EU icons to deepfakes and public-interest text, with a lighter approach for evidently artistic work.

4. Protect labels through export and editing

   Make sure labels persist when users download, export or edit content. Log any user requests in China for content without explicit labels, and keep those logs for at least six months.

5. Add detection and verification

   Offer a detection interface or tool for your marks. If you run a platform in China, check incoming content for implicit labels and flag declared or suspected AI content.

6. Document, test and monitor

   Record your labelling design in your AI governance documentation, for example under ISO 42001, test how robust labels are against compression and cropping, and track enforcement updates in both markets.

Common mistakes to avoid

• Relying on metadata alone, which is often stripped when content is re-uploaded or screenshotted
• Assuming the EU artistic exception means no disclosure at all
• Treating a quick human glance at AI-written news as "editorial responsibility"
• Forgetting that China's rules also bind platforms, app stores and users, not just generators
• Waiting for EU technical standards before marking outputs

A prompt to map your dual-market labelling obligations

Paste in your product details to get a first-draft compliance map. Your legal counsel in each jurisdiction should review the result.

You are an AI regulatory specialist familiar with EU AI Act
Article 50, the EU Code of Practice on marking and labelling
AI-generated content, China's Measures for Labeling AI-Generated
Synthetic Content and GB 45438-2025.

Our product:
[what it does; content types generated (text, image, audio,
video, virtual scenes); whether we generate, distribute or host
content; markets served; user types; whether content may depict
real people or events; current watermarking or metadata used]

Produce:
1. Our role(s) under each regime: EU provider and/or deployer;
   China generation service provider, content distribution
   platform, app distributor.
2. A table of obligations per content type for the EU and China:
   machine-readable marking, visible labels, exceptions.
3. Where one technical design can satisfy both regimes, and
   where we need market-specific behaviour.
4. Gaps in our current setup, ranked by legal risk.
5. Any deadlines that apply to us, including 2 December 2026.
6. Questions to put to local counsel in each market.

Use only what I provide and mark uncertain points UNKNOWN.

China labels the content so the ecosystem can police it. The EU marks the content so people can trust it. Global providers will need to do both, and do them well. — Kumail Mehdi

Frequently asked questions

Working on AI content labelling across the EU and China? Let's connect on LinkedIn. I'm always glad to compare notes on AI transparency.

Sources

• China Law Translate — Measures for Labeling AI-Generated Synthetic Content (translation)
• Covington Inside Privacy — China releases new labeling requirements for AI-generated content
• Technology's Legal Edge — China's Measures for labelling AI-generated content
• Bird & Bird — New AI content labelling rules in China compared to the EU AI Act
• Xinhua — Qinglang AI campaign phase 1 results (6 July 2026)
• Xinhua — Qinglang AI campaign phase 2 (2 September 2026)
• AI Regulations — China labelling enforcement cases timeline
• European Commission — Code of Practice on marking and labelling AI-generated content
• European Commission — Signing the Code of Practice and grandfathering for Article 50(2)
• Bird & Bird — The final transparency Code of Practice
• Paul, Weiss — EU finalises transparency rules for AI-generated content
• Stephenson Harwood — Commission adopts guidelines on Article 50
• Greenberg Traurig — Deepfakes, chatbots and AI-generated text under the AI Act