Under Article 50(4), deployers of AI systems that generate or manipulate image, audio or video content constituting a deepfake must disclose that the content has been artificially generated or manipulated. Deployers of systems that generate text published to inform the public on matters of public interest must disclose that the text was artificially generated, subject to the exception where the text has undergone human review and a natural or legal person holds editorial responsibility. The disclosure must be clear, distinguishable, and given in the manner Article 50 provides.
What is a deepfake for Article 50(4)?
A deepfake is defined in the Act 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. The definition turns on resemblance and on the false appearance of authenticity, not on the technical method used to create the content. A deployer should assess, for each piece of content, whether it resembles a real person, object, place, entity or event, and whether a person seeing it would reasonably take it to be authentic. Content that is clearly fictional, stylised, or presented in a way that makes its artificial nature obvious may fall outside the definition, but the assessment should be recorded with the content and context relied on, rather than assumed from the production method alone.
How must a deepfake be disclosed?
The deployer must disclose that the content has been artificially generated or manipulated, in a clear and distinguishable manner, in a way that is accessible and, where the content is published, visible to the audience. The disclosure should be attached to the content itself, so that it travels with the content when it is shared, as far as the deployer can ensure, and should not be placed only in a separate policy or in metadata that the audience will not see. The wording should state plainly that the content was artificially generated or manipulated, without ambiguous language that could be read as a general production credit. The deployer should test the disclosure in the publication channel, check that it remains visible on different devices and after sharing, and record the design and the test for each channel used.
For the broader regulatory framework, see Transparency duties for AI.
What about AI-generated text on matters of public interest?
Deployers of AI systems that generate text published with the purpose of informing the public on matters of public interest must disclose that the text was artificially generated, in a clear and distinguishable manner, unless the exception in Article 50(4) applies. Matters of public interest include, in general terms, topics that affect the public, such as news, public affairs, and matters of social importance, and the deployer should assess the purpose of publication, not only the topic in isolation. Text generated for internal drafting, for private communication, or for a purpose other than informing the public should be assessed separately, with the purpose recorded. A deployer that publishes AI-generated text on public-interest matters without considering this paragraph risks missing a disclosure duty that is distinct from the deepfake duty.
When does the human review exception apply?
The text disclosure duty does not apply where the AI-generated text has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the text. Both conditions matter: human review or editorial control must actually have taken place, and editorial responsibility must be held by an identifiable person who is responsible for the publication. A nominal review, where a person approves text without reading or checking it, or a process where no one holds editorial responsibility, does not meet the exception. The deployer or publisher should record the review process, the reviewer, the editorial responsibility holder, and the date, for each publication that relies on the exception, so that the reliance can be checked. The exception should be applied publication by publication, not as a blanket claim for all text.
How are artistic and similar works treated?
Article 50(4) provides that the deepfake disclosure duty does not apply, in the same way, to content that is part of an evidently artistic, creative, satirical, fictional or analogous work or programme, subject to the conditions and safeguards in the article, including that the disclosure must still be made in an appropriate manner that does not hamper the display or enjoyment of the work, where the article so provides. The treatment depends on whether the work is evidently artistic or creative, and on the manner of disclosure that is appropriate to the work. A deployer relying on that treatment should record the nature of the work, why it is evidently artistic or creative, and how disclosure was made in a manner appropriate to the work. A work that is presented as factual or news content, even where it uses creative techniques, should not be treated as artistic for this purpose without a careful assessment.
Continue with Exceptions in Article 50.
What does this guide not decide?
This guide does not decide whether particular content is a deepfake, whether text concerns a matter of public interest, whether human review was sufficient, or whether a work is evidently artistic. It does not assess a disclosure design or a publication. Those questions depend on the content, its resemblance and context, the publication purpose, the review process, and the consolidated legal text. A deployer or publisher should assess each publication on its facts, record the assessment and the disclosure, and confirm the position with a competent authority or legal adviser where the content is borderline or the exception is relied on.
Frequently asked questions
Must every AI-generated image be labelled a deepfake?
No. Article 50(4) deepfake disclosure applies to content that meets the deepfake definition: AI-generated or manipulated content resembling real persons, objects, places, entities or events that would falsely appear authentic. Other synthetic content is covered by the Article 50(2) marking duty for providers, and each piece of content should be assessed against the definition on its facts.
Does human editing remove the text disclosure duty?
Only where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication, as Article 50(4) provides. Both conditions must be met and recorded for the publication, and a nominal review without editorial responsibility does not meet the exception.
Are satirical works exempt from disclosure?
Not entirely. Article 50(4) provides a distinct treatment for evidently artistic, creative, satirical or fictional works, with disclosure in an appropriate manner that does not hamper enjoyment of the work, as the article provides. The work's nature and the manner of disclosure should be recorded, and factual or news content should not be treated as artistic without careful assessment.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk Article 50 ↗European Commission AI Act overview ↗