EU AI Content Transparency: Machine Marking and Human Disclosure
For EU-facing AI content, distinguish provider marking duties from publisher disclosure duties. Check content type, applicable scope and review exceptions against formal legal sources. Voluntary guidance alone does not establish compliance.
The applicable node has arrived, what should companies change?
The relevant transparency obligations under Article 50 of the EU Artificial Intelligence Act will apply from August 2, 2026. For marketing teams using generative AI for the EU market, the key change is to connect the way content is generated with publishing responsibilities, rather than adding the same conspicuous watermark to all AI-assisted materials.
The European Commission's note distinguishes between providers of systems that generate them and deployers that use them, and addresses technical marking and public disclosure respectively. Enterprises can design the release process accordingly, but whether it falls within the specific terms still needs to be verified based on the actual role, content and geographical scope of application. Judgments oriented to the EU cannot be directly extended to unified rules issued by all the world.
Position the role first, then determine the content to publish
If an enterprise only uses generation tools to produce website or social media materials, it should first check the publishing obligations of the deployer; if it also provides the generation system to others, it needs to separately identify the provider's responsibilities. The same enterprise may be involved in different activities and cannot be judged just by the name of "foreign trade company" or "marketing team".
Content classification determines the next step. General product introductions, simulation displays, and articles on public issues facing the public should not be combined just because they all use AI. The branches below help organize the facts. The actions in the table are recommendations for the release process and do not replace legal determinations.
| Actual activity or content | Responsibilities to check first | Next judgment and action |
|---|---|---|
| Provide generative AI systems | Machine-readable markup of relevant output | Find out how system output supports tagging and identification; ask technical support for confirmation if unclear. |
| Use AI to create images, audio or videos that may be passed off as real | Whether it constitutes deep forgery and corresponding disclosure | Check whether the content misleads people into believing it is the real object or event; if there is any doubt, refer it to professionals for judgment and disclosure arrangements. |
| Publish AI-generated or manipulated text involving public interest | Deployer-specific text disclosure requirements | Review the purpose of publication and the nature of the content; when claiming an exception, check the facts of human review and editorial responsibility. |
| Other AI-assisted marketing content | Do specific terms apply? | The generation purpose and release region data are retained; it is not directly concluded that the same label must be used based on the participation of AI. |
What problems do machine tagging and audience-facing instructions each solve?
Machine-readable tags serve the purpose of technical identification, enabling the relevant output to be identified as artificially generated or manipulated. It is not the same as the text description that readers see on the page. When enterprises learn about the tagging scheme from the tool provider, they should ask what type of output the tag is attached to, how to identify it, and whether relevant information is retained during downloading, editing, and publishing. The specific implementation method should be subject to actual tools and formal requirements.
Public disclosures are intended for people who view or read the content. Deep fakes involve pictures, audios or videos that make existing people, objects, places, entities or events look real, but may actually mislead the audience; public interest texts have corresponding text disclosure judgments. Product pictures cannot be automatically excluded from verification just because they are used for advertising, nor can they be deemed to be deep fakes simply because “the pictures are produced by AI”.
Before publishing, you can verify the document and the presentation page separately: on the document side, check whether the technical markup is still identifiable, and on the page side, check whether the necessary instructions are actually presented to the audience. If export or platform processing has changed the material, this step should be rechecked. Adding visible instructions does not prove that the machine mark has been obtained; nor does retaining technical information directly replace public disclosure where applicable.
Exceptions to manual review, subject to actual review and editorial responsibility
For AI-generated or manipulated texts that are oriented to the public and involve public interests, the Commission's note mentions a specific exception: the text has been manually reviewed and the editing responsibility is assumed by a natural or legal person. This involves both the review process and the attribution of responsibilities. "Someone has opened it and viewed it" is not enough to indicate that these conditions have been met.
Teams planning to publish accordingly can keep original and revised drafts, fact-checking records, review decisions, and information about the entities with editorial responsibility. The information should reflect what the person actually did, not fill in a signature form after posting. If only the tool automatically checks, or it is impossible to explain who is responsible for the editing results, you should first complete the facts and then ask a professional to confirm whether the exception applies.
This textual exception should not be used as a general exemption for deepfake images or videos. When publishing in a mixed manner, the text and audio-visual content should be judged separately to avoid treating the accompanying pictures and videos as having completed the disclosure review after the article has been reviewed by others.
Integrate generation, review and post-launch verification into a process
Foreign trade websites, social media and advertising channels can share content records, but the final presentation must be verified by channel. Just because the tool output conforms to a certain technical arrangement does not mean that the version uploaded, cropped, or forwarded by the marketing team retains the same information. Reviews should be of files and pages that are intended for actual publication.
The following checks each solve a problem. If there is a dispute about the classification, first supplement the content and purpose information; if the technical information is lost, go back to the export and platform processing links; if the person responsible is unknown, clarify the editing arrangements. Review the final version after the issue is resolved, and don’t treat “reviewed” as a substitute answer for everything.
- Record provenance: Save the build system used, the original output, and the manual modification process so that released versions can be traced back to production records.
- Identify the purpose: indicate whether it is a product display, simulated expression, or public issue information, and record the region and audience it is intended for.
- Verify supply support: Obtain output tags and identification instructions from the tool party, and distinguish supply commitments from documents actually obtained by the company.
- Implement manual responsibilities: record who checked what, modified what, and who assumes corresponding editorial responsibilities.
- Verify launch presentation: Check the information that needs to be audience-facing on the final page or platform preview to avoid description being cropped or hidden.
- Keep version evidence: Archive the online materials, review conclusions and actual presentation records; re-judge the affected matters when replacing materials later.
Voluntary Guidance Does Not Replace Legal Review
The accompanying Code provides a practical framework for providers and deployers to implement their transparency obligations and can be used to help demonstrate compliance. Adoption of the Code is voluntary, but the legal obligations associated with Article 50 are not themselves voluntary. Signing or adopting the Code is not a substitute for actual implementation, nor does it mean that a publication automatically meets all requirements.
When companies choose the guideline route, they should map the measures to their own tools, people, and release channels; when using other methods, they also need to be able to explain how the measures meet applicable requirements. Don’t confuse willingness to participate, vendor statements, and specific release evidence into the same conclusion.
Prepare authentic content samples, original materials, generation and modification records, review responsibility statements and release regional materials, and then ask legal professionals to confirm the scope of application based on the official text and the implementation status of member states. These inputs can help determine what needs to be disclosed, what technical aspects need to be substantiated, and whether there are facts to support the assertion of textual exceptions. Individual case liability and enforcement consequences cannot be directly presumed from process recommendations.
Common Problems
What Should You Confirm with an External Content Agency?
First find out whether the agency uses existing tools or provides a generation system, and obtains the generation source, modification records and disclosure arrangements of the delivered materials. The entrustment contract should clearly state who will provide evidence and who will perform release verification, but the division of labor in the contract cannot alone determine legal roles or exempt applicable responsibilities.
How should we troubleshoot old materials that have already been released online?
You can prioritize your searches by whether they are still being published, whether the original files can be found, and whether they could easily be mistaken for genuine content. Keep version records when redistributing or replacing. The specific application of historical materials should be verified individually, and all exemptions or violations cannot be asserted solely based on the original production time.
Is an "Illustration" Label Enough for AI Product Images?
This type of description helps explain the intended use, but it is not evidence of a technical mark and does not necessarily cover specific disclosure requirements. Also check that the overall presentation leads to the impression that the concept presentation is the actual product state, and that the specific descriptions are consistent with applicable requirements.
When publishing the same material to multiple markets, can a set of records be shared?
Generated sources and audit data can be used as a common manuscript, while regional applicable conclusions and final disclosure presentations should be recorded separately. This can not only reuse real evidence, but also avoid unconditionally applying the judgment of a certain market to other audiences.