Introduction: AI content labelling has become a common foundational action in China-EU AI governance

Since 2025, both China and the EU have entered an accelerated implementation phase for their regulatory systems on labelling AI-generated and synthesized content. On the Chinese side, the Cyberspace Administration of China and three other departments issued the Measures for Labelling AI-Generated and Synthesized Content, which clarify that AI-generated and synthesized content labels include explicit labels and implicit labels, and which took effect on 1 September 2025.[1] The supporting mandatory national standard GB 45438-2025, Cybersecurity Technology—Methods for Labelling AI-Generated and Synthesized Content, was also released on 28 February 2025 and took effect on 1 September 2025.[2]

On the EU side, Article 50 of the Artificial Intelligence Act (the "AI Act") sets out transparency obligations for certain AI systems, including that providers of AI systems mark synthetic audio, image, video or text outputs in a machine-readable format, and that deployers disclose deepfake content and certain text of public interest. In June 2026, the European Commission published the Code of Practice on Transparency of AI-Generated Content (the "EU Code of Practice") to support implementation of the obligations under Article 50(2), (4) and (5) of the AI Act regarding marking, detection and labelling of AI-generated content.[3]

On the surface, both the Chinese and EU rules point to the same proposition: "AI-generated content should be labelled." But in terms of institutional structure, the two sets of rules are not entirely the same. The Chinese rules take "service providers" as the main responsible entities, use "explicit labels + implicit labels" as the core structure, and implement the rules through dedicated measures and a mandatory national standard. The EU rules, by contrast, are layered under the AI Act between "providers" and "deployers," allocating "machine-readable marking," "detection" and "visible labelling" to different entities and different scenarios.

Therefore, for Chinese companies, the truly important question is not "whether the EU also requires AI content to be labelled," but rather: if a company has already established an AI-generated and synthesized content labelling mechanism in accordance with Chinese rules, what differentiated compliance requirements must it still satisfy when its products, services or content enter the EU market? This article takes the differences between China and EU AI content labelling rules as its main thread, compares them at the levels of entities, label types, applicable conditions, technical implementation and proof mechanisms, and proposes a "five patches" compliance path that Chinese companies can adopt when going global into the EU market: patch identity, patch scenarios, patch technology, patch labelling, patch evidence.

01. The basic structure of China's AI-generated and synthesized content labelling system

China's AI content labelling system is not a single rule, but a dedicated compliance requirement built on multiple institutional foundations, including algorithmic recommendation, deep synthesis, generative artificial intelligence and online information content governance. Article 1 of the Measures for Labelling AI-Generated and Synthesized Content expressly states that its basis for formulation includes the Cybersecurity Law, the Provisions on the Administration of Algorithmic Recommendations in Internet Information Services, the Provisions on the Administration of Deep Synthesis in Internet Information Services, the Interim Measures for the Administration of Generative Artificial Intelligence Services and other rules, and that its objectives are to promote the healthy development of artificial intelligence, regulate the labelling of AI-generated and synthesized content, protect the lawful rights and interests of citizens, legal persons and other organizations, and safeguard the public interest.[4] This shows that the institutional positioning of China's AI content labelling rules is not a mere consumer-notification rule, but one that combines online information content governance, technology regulation, platform responsibility and protection of the public interest.

As to applicable entities, Article 2 of the Measures for Labelling AI-Generated and Synthesized Content provides that internet information service providers that meet the circumstances under the relevant provisions on algorithmic recommendation, deep synthesis and generative artificial intelligence and that carry out AI-generated and synthesized content labelling activities are subject to these Measures; the Measures collectively refer to the relevant entities as "service providers."[5] This entity structure is consistent with China's existing internet information service regulatory system, that is, primary responsibility falls on the entities that provide the relevant internet information services to the public.

As to the scope of content, Article 3 of the Measures for Labelling AI-Generated and Synthesized Content provides that AI-generated and synthesized content means information such as text, images, audio, video and virtual scenes generated or synthesized using artificial intelligence technology.[6] This formulation covers the main forms of content production under current generative AI and deep synthesis technologies, and also provides the basis for subsequent technical standards to set labelling methods according to different content types.

As to label types, the Chinese rules expressly distinguish between "explicit labels" and "implicit labels." Explicit labels mean labels added to generated or synthesized content or interactive scenario interfaces, presented in the form of text, sound, graphics or other means, and clearly perceivable by users; implicit labels mean labels added to the file data of generated or synthesized content by technical measures and not easily perceivable by users.[7]

This structure reflects the two-tier governance approach of the Chinese rules: on the one hand, explicit labels enable ordinary users to directly perceive that content is AI-generated or synthesized; on the other hand, implicit labels provide a basis for platform governance, content tracing, technical detection and regulatory verification. In other words, the Chinese rules do not merely require adding a sentence such as "this content was generated by AI" to a page; they seek, by combining user-perceivable labels with technical labels, to reduce the risk that AI-generated content is mistaken for real content, maliciously disseminated or used to evade regulation.

Supporting the Measures for Labelling AI-Generated and Synthesized Content, GB 45438-2025, Cybersecurity Technology—Methods for Labelling AI-Generated and Synthesized Content, was released on 28 February 2025 and took effect on 1 September 2025. The National Public Service Platform for Standards Information shows that the standard is current, its category is "security," its competent department is the Office of the Central Cyberspace Affairs Commission, and its technical committee is the National Technical Committee 260 on Cybersecurity of Standardization Administration of China.[8] The significance of this standard is that it moves explicit labels and implicit labels beyond principled requirements into an executable phase covering different content types, different technical methods and different file data structures.

Overall, China's AI content labelling system can be summarized as follows: centred on service providers, with explicit labels and implicit labels as the basic structure, with dedicated measures and national standards as the implementation path, and with preventing public confusion, maintaining online information order and promoting the healthy development of artificial intelligence as the main institutional objectives.

02. The basic structure of the EU's AI-generated content transparency rules: mandatory AI Act, voluntary Code

The institutional source of the EU's AI content labelling rules is mainly Article 50 of the AI Act. Unlike the Chinese rules, the AI Act places this under the framework of "transparency obligations for certain AI systems," emphasizing transparency arrangements to reduce the risks of deception, manipulation and harm to the information ecosystem.

Article 50(1) of the AI Act first provides that AI systems intended to interact directly with natural persons shall be designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a reasonably well-informed, observant and circumspect natural person, taking into account the circumstances and the context of use.[9] Article 50(2) further provides that providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated; the relevant technical solutions shall, to the extent technically feasible, be effective, interoperable, robust and reliable.[10]

Article 50(4) provides that deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake shall disclose that the content has been artificially generated or manipulated; where the content is part of an evidently artistic, creative, satirical, fictional or analogous work, the disclosure obligation may be limited to disclosing the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work. Deployers of an AI system that generates or manipulates text published for the purpose of informing the public on matters of public interest shall also disclose that the text has been artificially generated or manipulated; an exception may apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.[11] Article 50(5) requires that the above information be provided in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, and in compliance with applicable accessibility requirements.[12]

The EU Code of Practice published by the European Commission is built around the above Article 50(2), (4) and (5). The European Commission's official page expressly states that the Code aims to support the AI Act's transparency obligations on marking and labelling of AI-generated content; the Code is divided into two parts: the first part is addressed to providers and sets out marking and detection of AI-generated and manipulated content; the second part is addressed to deployers and sets out labelling of deepfakes and AI-generated or manipulated text.[13]

Particular attention should be paid to the fact that the EU Code of Practice is voluntary, but the transparency obligations under Article 50 of the AI Act are themselves legal obligations. The European Commission's official page expressly states that, although adherence to the Code is voluntary, the transparency requirements under Article 50 of the AI Act are legal obligations.[14] Therefore, a company may choose not to sign the EU Code of Practice and may choose not to use the icon scheme provided by the EU, but as long as it falls within the scope of application of Article 50 of the AI Act, it must still perform the corresponding transparency obligations and be prepared to demonstrate the adequacy of its alternative measures.

From a practical perspective, the value of the EU Code of Practice lies not in exempting liability, but in providing a compliance path with higher regulatory recognition, lower interpretation costs and greater certainty of application across Member States. The European Commission's official page further explains that, after signing the Code and receiving a positive assessment, providers and deployers may rely on the Code's measures to demonstrate compliance with the AI Act rules on labelling and detection of AI-generated content, deepfakes and publication of certain texts; entities that choose other means of compliance must demonstrate that their measures are adequate, and may be assessed case by case by different market surveillance authorities.[15]

In addition, the AI Act has clear extraterritorial application features. Article 2 of the AI Act provides that it applies to providers placing AI systems on the market or putting them into service in the EU, or placing general-purpose AI models on the EU market, regardless of whether they are established or located in the EU; it also applies to providers and deployers located in third countries, as long as the output produced by the AI system is used in the EU.[16] This means that Chinese companies, even if they have not established an entity in the EU, may still be brought within the scope of application of the AI Act as long as the output of their AI systems enters use scenarios in the EU.

Therefore, the EU rules on AI content identification can be summarized as follows: Article 50 of the AI Act serves as the statutory obligation basis, provider/deployer serves as the layered allocation of responsibility, marking, detection and labelling serve as the obligation structure, and the EU code of practice serves as a voluntary but evidentially valuable compliance pathway.

03. Subject differences: China's "service providers" and the EU's "Provider / Deployer"

The first key difference between the Chinese and EU rules on AI content identification is the different structure of responsible subjects.

The Chinese rules mainly revolve around "service providers." Article 2 of the Measures for the Identification of AI-Generated Synthetic Content clearly provides that network information service providers that meet the circumstances under the relevant provisions on algorithmic recommendation, deep synthesis, and generative AI shall be subject to these Measures when carrying out activities for the identification of AI-generated synthetic content.[17] In the Chinese context, companies usually first determine whether they constitute generative AI service providers, deep synthesis service providers, or related network information service providers, and then determine whether they need to perform obligations such as explicit identification, implicit identification, user prompts, and dissemination management.

The EU rules adopt a more detailed layering of subjects. Article 3 of the AI Act defines "provider" as a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model, or that has an AI system or a general-purpose AI model developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge.[18] "Deployer" means a natural or legal person, public authority, agency or other body using an AI system under its authority, but excluding personal non-professional use.[19]

In other words, in the EU context, a provider is closer to "the supplier of an AI system, model, tool or API"; a deployer is closer to "the party that uses an AI system in its own business and controls the purpose of its use or the publication scenario." This layering is particularly important for Chinese companies, because the same company may have different identities in different business scenarios.

For example, if a Chinese company develops an AI image generation tool and provides SaaS or API services to EU customers under its own brand, it will very likely constitute a provider under EU law and need to bear machine-readable marking and detection obligations. If the company also uses the tool to generate advertising images for its own brand and places them with EU consumers, then in that advertising publication scenario it may also constitute a deployer and need to determine whether deepfakes are involved and whether visible labelling is required.

For another example, if a company merely calls third-party AI tools to generate EU-market press releases, product descriptions, investor relations materials or public communication copy, under the Chinese rules this may mainly manifest as internal AI use management or content publication management issues; but under the EU rules, if the text is AI-generated or manipulated and is published for the purpose of informing the public on matters of public interest, the company may bear disclosure obligations as a deployer.

Therefore, the first step for Chinese companies entering the EU market is not to choose a label style, but to re-identify their role under the AI Act. Companies need to determine separately by product line, business scenario and control over content publication: whether they are a provider, a deployer, or simultaneously bear a dual role in different links. An incorrect determination of subject identity will lead to incorrect configuration of subsequent marking, detection and labelling obligations, and will also affect the allocation of contractual liability between the company and upstream and downstream customers.

04. Differences in identification types: China's "explicit/implicit identification" and the EU's "Labelling / Marking / Detection"

The second key difference between the Chinese and EU rules is the different types of identification and functions of obligations.

The Chinese rules clearly distinguish between explicit identification and implicit identification. Explicit identification is presented through text, sound, graphics or other means and can be clearly perceived by users; implicit identification is added to the data of generated synthetic content files through technical measures and is not easily perceived by users.[20] Functionally, explicit identification is aimed at ordinary users and is intended to reduce user misidentification; implicit identification is aimed at technical recognition, platform governance and regulatory tracing, and is intended to provide a basis for content source identification and technical regulation.

The EU rules split the transparency of AI-generated content into three functions: marking, detection and labelling. First, marking is a provider obligation. Article 50(2) of the AI Act requires that providers of AI systems generating synthetic audio, image, video or text content shall ensure that the outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated.[21] This means that marking is not a label directly seen by ordinary users, but a machine-readable technical mark embedded in or associated with the content.

Second, detection is the detection capability that accompanies marking. Article 50(2) of the AI Act provides in parallel that outputs are "marked in a machine-readable format" and "detectable as artificially generated or manipulated," indicating that the EU is not satisfied with companies claiming that they "have added a mark," but requires that the mark be detectable.[22] The EU code of practice also summarizes the first part addressed to providers as "marking and detection of AI-generated and manipulated content."[23]

Third, labelling is a deployer obligation. Article 50(4) of the AI Act requires that deployers using an AI system to generate or manipulate image, audio or video content constituting a deepfake shall disclose that the content is artificially generated or manipulated; when publishing AI-generated or manipulated text to inform the public on matters of public interest, they shall also disclose that fact.[24] The European Commission's press release also clearly states that deepfakes and AI-generated or AI-manipulated text published on matters of public interest must be clearly labelled.[25]

From the perspective of conceptual correspondence, China's "explicit identification" and the EU's labelling have a certain functional similarity, both being perceptible prompts aimed at ordinary users; China's "implicit identification" and the EU's marking also have a certain similarity, both emphasizing machine or system recognition through technical means. However, this correspondence cannot be simply equated.

First, the subjects cannot be equated. The Chinese rules mainly set explicit and implicit identification obligations around service providers, while the EU rules separately allocate providers' marking/detection obligations and deployers' labelling obligations. Second, the functions cannot be equated. China's implicit identification emphasizes adding technical identifiers to file data, while EU marking also requires machine-readability and detectability, and further emphasizes effectiveness, interoperability, robustness and reliability. Third, the scenarios cannot be equated. EU labelling specifically targets deepfakes and public-interest text; not all AI-generated content follows exactly the same disclosure logic on the deployer side. Finally, the proof cannot be equated. A company merely proving that "I added an implicit identifier" does not automatically prove that it has satisfied the marking and detection requirements under Article 50 of the EU AI Act.

Therefore, even if Chinese companies already have explicit and implicit identification, when entering the EU market they still need to fill in the division of responsibility, technical capabilities and proof mechanisms among marking, detection and labelling.

05. Differences in applicable conditions: which AI-generated content needs to be identified

The third difference between the Chinese and EU rules is the different applicable conditions and trigger scenarios.

The Chinese rules take AI-generated synthetic content as the basic object. Article 3 of the Measures for the Identification of AI-Generated Synthetic Content defines it as information such as text, images, audio, video and virtual scenes generated or synthesized using AI technology.[26] From the logic of the rules, the Chinese system pays more attention to the identification and management of AI-generated synthetic content in the links of generation, display, dissemination, download and copying, with the focus on preventing the public from misidentifying the source, authenticity and attributes of content.

The EU rules set trigger conditions separately for providers and deployers. For providers, as long as their AI systems, including general-purpose AI systems, generate synthetic audio, image, video or text content, they need to ensure that the outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated. Article 50(2) of the AI Act also provides that this obligation applies to the extent technically feasible, and may take into account the characteristics of the content type, implementation costs and the generally acknowledged state of the art.[27] In addition, the AI Act provides exceptions for AI systems that only perform a standard editing assistance function and do not substantially alter the input data provided by the deployer or their semantics, and for scenarios used by law for the detection, prevention, investigation or prosecution of criminal offences.[28]

For deployers, Article 50(4) of the AI Act sets more specific trigger conditions. The first category is the use of an AI system to generate or manipulate image, audio or video content constituting 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 to a person to be authentic or truthful.[29] The second category is the use of an AI system to generate or manipulate text and publish it for the purpose of informing the public on matters of public interest. Such text needs to disclose its AI-generated or manipulated nature, but an exception may apply if it has undergone human review or editorial control and a natural or legal person holds editorial responsibility.[30]

This gives EU rules a distinctly scenario-based character on the deployer side. Companies cannot simply ask "is the content AI-generated"; they must further determine: does the content constitute a deepfake? Is it used to inform the public about matters of public interest? Might users mistake it for real? Has it undergone genuine, effective human review or editorial control? Is the disclosure made at the point of the user's first exposure to the content?

For Chinese companies, these determinations are especially important. For example, if an AI-generated marketing poster uses fictional characters and would not lead users to mistake them for specific real persons, whether it constitutes a deepfake must be determined in light of the specific content; AI-generated brand promotional copy is generally not necessarily a public-interest text, but where it involves financial investment, public health, public safety, major social events, policy interpretation, ESG, or statements on major public events of an enterprise, the likelihood of triggering the public-interest text disclosure obligation increases significantly. As another example, if an AI-assisted news release has undergone substantive human editorial control and the company or media organization bears editorial responsibility, there may be room to apply the EU text disclosure exception, but the company needs to retain the corresponding review records and evidence of responsibility assumption.

06. Technical differences and the "five patches" for companies going global: how to connect with EU requirements on the basis of Chinese compliance

All the aforementioned differences must ultimately come down to how companies rectify their practices. For companies that have already established AI content labeling mechanisms in accordance with Chinese rules, EU compliance should not start from scratch. A more reasonable path is, on the basis of the existing Chinese "explicit labeling + implicit labeling + content review + technical standards conformity" framework, to conduct a China-EU gap assessment and supplement the EU-specific modules around Article 50 of the AI Act.

This article proposes to summarize the AI content labeling compliance path for Chinese companies' AI going global to the EU as "five patches": patch identity, patch scenarios, patch technology, patch labeling, and patch evidence.

(1) Patch identity: re-identify the Provider / Deployer role

"Patching identity" is the first step of EU compliance. Companies need to re-identify their role under the EU AI Act by product line and business scenario, rather than judging solely on the basis of "service provider" under Chinese rules.

Companies should at least sort out the following questions: Do they develop AI systems or general-purpose AI models? Do they provide AI systems to the EU market under their own name or trademark? Do they provide generative AI capabilities to EU customers through APIs, SaaS, SDKs, or embedded products? Do they merely use third-party AI systems to generate content? Do they themselves control content publication? Is the output produced by the AI system used in the EU?

In terms of practical documents, companies can create an "EU AI Act Provider / Deployer Role Identification Table," listing item by item "product/business, AI function, target market, corporate conduct, EU status, corresponding obligations." For example, when an AI image generation tool is opened to EU users, the company may bear marking and detection obligations as a provider; when a company uses AI tools to generate EU advertising materials and publishes them itself, it may bear labeling obligations as a deployer; when an AI digital human service company both provides generation tools and participates in content delivery and publication control, it needs to determine whether it has both provider and deployer status.

(2) Patch scenarios: supplement deepfake, public-interest text, and first-exposure determinations

"Patching scenarios" is the core of the EU deployer obligations. Chinese companies in domestic compliance usually already focus on content types such as text, images, audio, video, and virtual scenes, but after entering the EU, they need to add three scenario determinations.

The first is the deepfake determination. Companies need to determine whether the relevant AI-generated or manipulated content is image, audio, or video, whether it resembles real persons, objects, places, entities, or events, and whether it may lead users to mistake it for real or trustworthy. Especially in scenarios such as digital humans, virtual anchors, AI face-swapping, AI voiceover, AI restoration of historical figures, AI-generated news scene footage, and AI simulation of real products or real events, stricter review checkpoints should be established.

The second is the public-interest text determination. When publishing AI-generated or manipulated text, companies should determine whether it is used to inform the public about matters of public interest. For scenarios such as news and information, public health, financial investment, policies and regulations, public safety, elections and politics, major social events, ESG report summaries, and statements on major public events of enterprises, companies should carefully assess whether the disclosure obligation is triggered.

The third is the first-exposure determination. Article 50(5) of the AI Act requires that the relevant information be provided at the latest at the time of the first interaction or first exposure.[31] Therefore, companies cannot disclose only on content detail pages or in terms of service; they should identify the actual point at which users first encounter the content, such as the beginning of a video, the start of audio playback, the image display area, social media preview pages, advertising landing pages, app push cards, search result summaries, etc., and provide clear, discernible notice at that point.

(3) Patch technology: upgrade from implicit labeling to Marking / Detection capabilities

"Patching technology" is the most easily underestimated part of the EU provider obligations. If Chinese companies have already added implicit labels, watermarks, or metadata in accordance with Chinese rules, that does not automatically mean they satisfy the EU marking obligation. Article 50(2) of the EU AI Act requires that outputs be marked in a machine-readable format and be detectable as artificially generated or manipulated; the relevant technical solutions should be effective, interoperable, robust, and reliable to the extent technically feasible.[32]

From the perspective of corporate implementation, the EU technical gap-filling can be summarized as the "four ables": readable, detectable, retainable, and provable. "Readable" means that the markings in AI-generated content should be machine-readable, not merely recognizable by the human eye. "Detectable" means that companies should provide detection paths, detection tools, or detection interfaces so that deployers, platforms, regulators, or other relevant parties can verify whether content was generated or manipulated by their AI systems. "Retainable" means that markings should, as far as possible, be retained during common dissemination processes such as downloading, copying, transcoding, compression, screenshots, editing, and format conversion, or that the risk of loss be reduced through multi-layered mechanisms. "Provable" means that companies can prove through test reports, technical descriptions, detection records, version logs, etc., that their technical solutions have reached a reasonable level.

The EU code of practice emphasizes marking and detection in the provider section and treats it as the core content of Part I of the code.[33] Although the technical implementation paths for different content types will still be limited by technological maturity, companies should at least establish a technical assessment mechanism to determine whether existing solutions such as watermarks, metadata, digital signatures, detection interfaces, and free-text watermarking can meet the EU rules' requirements for machine readability, detectability, and provability.

In terms of practical documents, companies can produce materials such as the "AI-Generated Content Marking Technical Solution Description," "AI-Generated Content Detection Interface Description," "Watermark and Metadata Retention Test Report," and "Free-Text Marking Exception and Alternative Measures Description." These documents not only serve internal compliance but can also be used for customer due diligence, platform access, regulatory communication, and dispute response.

(4) Patch labeling: establish Deployer user-facing Labelling display rules

"Patching labeling" is the step that translates technical compliance into user-perceivable transparency. For deployers, the compliance focus is not to prove how the underlying model is marked, but to ensure that when users first encounter deepfake content or public-interest text, they can clearly and promptly understand its AI-generated or manipulated nature.

The European Commission's press release clearly states that, from the date the relevant AI Act obligations apply, deepfakes and AI-generated or AI-manipulated text published on matters of public interest must be clearly labeled; users should also be informed when interacting with interactive AI systems such as chatbots.[34] In conjunction with Article 50(5) of the AI Act, labeling should also be clear, discernible, and comply with applicable accessibility requirements.[35]

Companies should develop display specifications by content type. Images can have an AI icon or text placed in a position that is not easily obscured; videos should display it at the beginning and, where necessary, repeat it periodically and redisplay it after interruption; pure audio should include a brief audio statement at the beginning; text can be labeled near the title, at the beginning of the body, or at the relevant paragraph; where only part of the content is AI-generated or modified, the corresponding part can be labeled. For artistic, creative, satirical, or fictional works, a light-touch disclosure method that does not impede the display and appreciation of the work may be used.

The AI icon provided by the EU can be used as an optional tool, but it is not the only compliant method. For Chinese companies lacking experience in EU-localized label design, using the EU icon or referring to its design logic may help reduce user comprehension costs and regulatory communication costs. However, companies should still note that using a particular icon does not automatically equal compliance; the key remains whether the label is provided in the right scenario, at the right time, and in the right position, in a clear, discernible, and understandable manner.

(5) Patch evidence: establish contracts, logs, exceptions, and supply-chain responsibility allocation

"Patching evidence" is the key for AI content labeling compliance to move from "done internally" to "externally provable." Especially in the EU market, companies must not only be able to explain that they have a labeling system, but also prove how obligations are allocated between upstream and downstream, how technical measures are implemented, when content is labeled, and why exceptions apply.

Companies should establish at least four types of evidence. The first is technical evidence, including watermark test reports, metadata retention records, detection interface descriptions, detection result samples, system version records, and assessments of the effectiveness and robustness of technical solutions. The second is content evidence, including AI generation records, human editing records, publication approval records, records of determinations on whether content constitutes a deepfake, records of determinations on whether it is public-interest text, screenshots of labeling styles, and archived records of first-exposure pages or playback interfaces.

The third category is contractual evidence, including clauses on whether upstream AI tools support EU marking and detection, clauses on whether downstream customers bear labelling obligations, clauses prohibiting the removal or tampering of watermarks or metadata, detection assistance clauses, regulatory investigation assistance clauses, and clauses allocating liability for marking failure, incorrect marking, or removal of markings. The fourth category is exception evidence, including assessments of the artistic creation exception, records of human review and editorial control, records of the entity responsible for editing, information on the person responsible for publication, and records of internal review processes. In particular, enterprises wishing to rely on the EU public interest text disclosure exception should demonstrate that AI-generated text has indeed undergone substantive human review or editorial control and that a natural or legal person bears editorial responsibility, and cannot claim exemption merely on the ground that "a human looked at it."

It can thus be seen that for Chinese enterprises taking AI overseas to the EU, the task is not to translate domestic AI labels into English, nor simply to add an "AI" icon, but to supplement the domestic "explicit marking + implicit marking" with five EU compliance modules: "identity, scenario, technology, labelling, and evidence."

This can be further summarized as: China lays the foundation, the EU adds five items—identity must be re-determined, scenarios must be subdivided, technology must be detectable, labelling must be visible, and evidence must be provable. This approach does not require enterprises to start over, but rather converts China's existing marking compliance system into a modular upgrade plan that can connect with Article 50 of the EU AI Act.

Conclusion: The key to overseas compliance for AI content marking is not "sticking on labels" but "building mechanisms"

Both China and the EU have incorporated AI-generated content marking into their statutory compliance frameworks, but the institutional logic of the two sets of rules is not the same. China's rules take service providers, explicit marking, implicit marking, and national standards as their main line; the EU rules are based on Article 50 of the AI Act, distinguish between provider and deployer, and separately set marking, detection, and labelling obligations.

For Chinese enterprises, the key to overseas compliance for AI content marking is not to build a completely independent EU system from scratch, but to carry out gap assessments and modular upgrades on the basis of existing Chinese compliance. Enterprises should first determine their subject identity under EU law, then identify EU-specific scenarios such as deepfakes, public interest texts, and first contact; on the technical side, upgrade domestic implicit marking capabilities into machine-readable, detectable, retainable, and provable marking and detection mechanisms; on the content publishing side, establish labelling display rules oriented toward natural persons; and on the management side, supplement contracts, logs, approvals, exception assessments, and supply chain liability allocation.

In the future, AI content marking will no longer be a single notice on a page, but a mechanism composed jointly of enterprise AI governance, content compliance, technical architecture, and cross-border business management. Whoever can complete the compliance connection between Chinese rules and EU rules earlier will be able to enter the EU AI product, service, and content markets at lower cost.

Endnotes

Questions and answers

What is the difference between China's explicit and implicit labels for AI-generated content?

Explicit labels are added to content or interactive interfaces in text, sound, graphics or other forms clearly perceivable by users. Implicit labels are added to file data by technical measures and are not easily perceivable. Together they support user awareness and platform governance, tracing and regulatory verification.

Who bears the labelling obligations under Article 50 of the EU AI Act?

Providers of AI systems generating synthetic audio, image, video or text must mark outputs in a machine-readable format and make them detectable. Deployers must disclose deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest, at the latest at first interaction or exposure.

Is the EU Code of Practice on Transparency of AI-Generated Content mandatory?

No. Adherence to the Code is voluntary, but the transparency obligations under Article 50 of the AI Act are legal obligations. Companies may choose other means but must demonstrate their measures are adequate, and may be assessed case by case by market surveillance authorities.

Does the EU AI Act apply to Chinese companies without an EU establishment?

Yes. Article 2 of the AI Act applies to providers placing AI systems on the EU market or putting them into service in the EU, and to providers and deployers in third countries where the output produced by the AI system is used in the EU.

What are the 'five patches' for Chinese companies to align with EU AI content labelling rules?

The five patches are: patch identity (re-determine provider/deployer role), patch scenarios (deepfake, public-interest text, first exposure), patch technology (machine-readable, detectable, retainable, provable marking), patch labelling (user-facing display rules), and patch evidence (technical, content, contractual and exception records).

Laws and cases cited

Laws

  • Cybersecurity Law

Regulations

  • Measures for Labelling AI-Generated and Synthesized Content
  • Artificial Intelligence Act
  • Provisions on the Administration of Algorithmic Recommendations in Internet Information Services
  • Provisions on the Administration of Deep Synthesis in Internet Information Services
  • Interim Measures for the Administration of Generative Artificial Intelligence Services

Standards

  • GB 45438-2025, Cybersecurity Technology—Methods for Labelling AI-Generated and Synthesized Content

Other

  • Code of Practice on Transparency of AI-Generated Content

References

  1. Cyberspace Administration of China and three other departments: Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content, issued on March 14, 2025, effective September 1, 2025; the Measures specify that labels for AI-generated synthetic content include explicit labels and implicit labels. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  2. National Standards Information Public Service Platform: GB 45438-2025 Cybersecurity Technology—Labeling Method for Content Generated by Artificial Intelligence, release date February 28, 2025, implementation date September 1, 2025, standard status current, standard category security. Source: https://std.samr.gov.cn/gb/search/gbDetailed?id=301E0388CB75788DE06397BE0A0AE1B4
  3. European Commission, Code of Practice on Transparency of AI-Generated Content. The European Commission page states that the Code supports the transparency obligations of the AI Act on marking and labelling of AI-generated content, that the applicable obligations take effect from August 2, 2026, and that it covers Article 50(2), (4), and (5) of the AI Act. Source: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
  4. Article 1 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  5. Article 2 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  6. Article 3, paragraph 1 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  7. Article 3, paragraphs 2 to 4 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  8. National Standards Information Public Service Platform: basic information on GB 45438-2025 Cybersecurity Technology—Labeling Method for Content Generated by Artificial Intelligence. Source: https://std.samr.gov.cn/gb/search/gbDetailed?id=301E0388CB75788DE06397BE0A0AE1B4
  9. Regulation (EU) 2024/1689, Article 50(1). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  10. Regulation (EU) 2024/1689, Article 50(2). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  11. Regulation (EU) 2024/1689, Article 50(4). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  12. Regulation (EU) 2024/1689, Article 50(5). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  13. European Commission, Code of Practice on Transparency of AI-Generated Content. The page states that the Code is divided into Section 1 Providers - Rules for marking and detection of AI-generated and manipulated content; Section 2 Deployers - Rules for labelling of deepfakes and AI-generated and manipulated text. Source: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
  14. European Commission, Code of Practice on Transparency of AI-Generated Content. The page states that adherence to the code is voluntary, the transparency requirements under Article 50 of the AI Act are legal obligations. Source: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
  15. European Commission, Code of Practice on Transparency of AI-Generated Content. The page states that after a positive assessment, signatories may rely on the Code's measures to demonstrate compliance; entities that do not achieve compliance through the Code must demonstrate that their measures are adequate and are subject to case-by-case assessment. Source: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
  16. Regulation (EU) 2024/1689, Article 2(1)(a) and Article 2(1)(c). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  17. Article 2 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  18. Regulation (EU) 2024/1689, Article 3(3), definition of “provider”. Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  19. Regulation (EU) 2024/1689, Article 3(4), definition of “deployer”. Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  20. Article 3, paragraphs 2 to 4 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  21. Regulation (EU) 2024/1689, Article 50(2). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  22. Regulation (EU) 2024/1689, Article 50(2). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  23. European Commission, Code of Practice on Transparency of AI-Generated Content, Section 1: Providers - Rules for marking and detection of AI-generated and manipulated content. Source: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
  24. Regulation (EU) 2024/1689, Article 50(4). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  25. European Commission, Commission publishes Code of Practice on marking and labelling AI-generated content. The press release states that deepfakes and AI-generated or AI-manipulated text published on matters of public interest must be clearly labelled. Source: https://digital-strategy.ec.europa.eu/en/news/commission-publishes-code-practice-marking-and-labelling-ai-generated-content
  26. Article 3, paragraph 1 of the Measures for the Labeling of Artificial Intelligence-Generated Synthetic Content. Source: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  27. Regulation (EU) 2024/1689, Article 50(2). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  28. Regulation (EU) 2024/1689, Article 50(2), final sentence. Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  29. Regulation (EU) 2024/1689, Article 3(60), definition of “deep fake”; Article 50(4) sets the disclosure obligation for deep fake content. Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  30. Regulation (EU) 2024/1689, Article 50(4). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  31. Regulation (EU) 2024/1689, Article 50(5). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  32. Regulation (EU) 2024/1689, Article 50(2). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  33. European Commission, Code of Practice on Transparency of AI-Generated Content. Source: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
  34. European Commission, Commission publishes Code of Practice on marking and labelling AI-generated content. Source: https://digital-strategy.ec.europa.eu/en/news/commission-publishes-code-practice-marking-and-labelling-ai-generated-content
  35. Regulation (EU) 2024/1689, Article 50(5). Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng