Do You Need to Label AI-Generated Content?
The EU AI Act does not require a blanket visible label on all AI content. Article 50 distinguishes between technical marking by providers, deepfakes, and AI text on matters of public interest. A practical overview — not legal advice.
No. The EU AI Act does not impose a blanket requirement to place a visible "AI-generated" label on every piece of content created with artificial intelligence. The rules depend on the type of content, how it is used and whether you are the provider of the AI system or the organisation deploying it. Deepfakes and certain AI-generated texts dealing with matters of public interest are particularly relevant.
Providers of AI systems that generate synthetic text, images, audio or video have a separate obligation: their systems must, in principle, make generated or manipulated output machine-readable and detectable as such. That technical marking requirement is different from placing a visible warning underneath every AI-generated image. Article 50 of the EU AI Act therefore creates several transparency duties rather than one universal labelling rule — for the broader picture of what else already applies, see our overview The EU AI Act for SMEs.
What does Article 50 of the AI Act require?
Article 50 draws an important distinction between the provider of an AI system and the organisation or individual deploying it.
Under Article 50(2), providers of AI systems — including general-purpose AI systems — that generate synthetic audio, images, video or text must ensure that their output is marked in a machine-readable format and can be detected as artificially generated or manipulated. The technical solution should be effective, interoperable, robust and reliable as far as technically feasible, taking into account the characteristics of the content, implementation costs and the state of the art.
There is an exception where the AI merely provides an assistive function for standard editing or does not substantially alter the input or its meaning. A minor AI-based correction is therefore not necessarily treated in the same way as generating an entirely new photograph, voice recording or article.
Different obligations apply to deployers who actually use and publish the resulting content.
Images, audio and video: deepfakes are the key category
Under the AI Act, not every AI-generated image is automatically a deepfake.
The Regulation defines a deepfake as AI-generated or manipulated image, audio or video content that resembles existing people, objects, places, entities or events and could falsely appear authentic or truthful.
Article 50(4) requires deployers using AI to generate or manipulate content constituting a deepfake to disclose that the material has been artificially generated or manipulated. Examples include:
- a realistic video in which an existing executive appears to make a statement they never made;
- an AI-cloned voice imitating a real person;
- a convincing image showing a real-looking event that never happened.
There is a more proportionate rule for clearly artistic, creative, satirical or fictional works. Disclosure is still relevant, but it may be made in a way that does not unnecessarily interfere with the display or enjoyment of the work.
Does AI-generated text need a label?
Again, not automatically.
Article 50(4) specifically addresses AI-generated or manipulated text published for the purpose of informing the public about matters of public interest. In those circumstances, its artificial origin must in principle be disclosed.
There is an important exception. The disclosure obligation does not apply where the AI-generated content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication.
This distinction matters for ordinary business publishing.
A company may use an AI assistant to produce the first draft of an article, after which an editor verifies sources, rewrites sections, removes errors and accepts responsibility for the final publication. That is very different from an automated system publishing large volumes of news-like content with no meaningful human review.
The relevant question is therefore not simply, "Was AI involved?" The purpose of the publication, the extent of editorial control and responsibility for the final output all matter.
Legal requirements and editorial policy are not the same thing
An organisation can choose to disclose AI use even where Article 50 does not require a specific visible label.
For example, a website may publish a general editorial statement explaining that AI is used to assist with research, translation, summarisation or image creation while human editors remain responsible for published material.
Such a policy can make sense for reasons of trust, governance, professional standards or sector-specific expectations. It should not, however, be presented as evidence that the AI Act requires every piece of AI-assisted content to carry a visible warning.
Likewise, a voluntary generic disclaimer does not necessarily satisfy a specific legal disclosure obligation when one applies.
Metadata, watermarks and provenance
A visible label is only one transparency mechanism. A practical technical stack can combine:
- machine-readable metadata recording AI involvement;
- visible or invisible watermarks;
- cryptographically signed provenance information;
- a human-readable disclosure alongside the content;
- provenance records showing how a media file was created and edited.
C2PA is one example of a provenance framework. It provides a technical standard for attaching verifiable information about the origin and history of digital content. It is not itself an AI Act requirement, and a C2PA credential does not prove that the factual message contained in an image or video is true. It provides evidence about provenance and recorded transformations.
A robust system should not depend entirely on one mechanism. Metadata can disappear when content is exported or uploaded to another service, while a visible label can be cropped from an image. Machine-readable marking, provenance records and human-readable disclosure therefore serve different purposes and can reinforce each other.
When is a visible disclosure a sensible choice?
A visible disclosure is particularly useful whenever a reasonable viewer might otherwise believe that synthetic material is authentic. Examples include:
- "AI-generated image."
- "This video contains AI-generated material."
- "The voice in this recording is synthetic."
For an obviously fictional illustration, a highly prominent warning may be less useful. The position changes when the content depicts realistic people, real-world events, political or social issues, news-like material or something that could be interpreted as documentary evidence.
For the relevant Article 50 duties, disclosure must be clear and distinguishable and provided no later than the person's first interaction with or exposure to the content. The European Commission's current guidance also makes clear that, in the case of deepfakes, deployers cannot simply rely on an invisible machine-readable marker placed in the file by the AI provider.
What are the risks of not labelling?
The most obvious risk is regulatory. Where Article 50 requires disclosure and an organisation fails to provide it, that can result in enforcement under the AI Act.
There are also operational and reputational risks. Realistic synthetic content published without adequate context can mislead users and damage trust in other material published by the same organisation. Provenance can become important during a dispute as well: being able to show how a piece of content was generated, what modifications were made and who approved the final version may be considerably more useful than relying only on a label added at publication time.
Public bodies, legal organisations and companies communicating about sensitive public issues should therefore treat labelling as part of a wider content-governance process that also covers provenance, review and accountability.
Frequently asked questions
Does every AI-generated image have to carry a visible AI label under the EU AI Act? No. Article 50 does not create a universal visible-labelling requirement for every AI image. Providers have technical marking obligations in relevant cases, while deployers have specific disclosure obligations, including for content that qualifies as a deepfake.
Does an article drafted with ChatGPT need to be labelled? Not automatically. The text-related rule in Article 50(4) concerns AI-generated or manipulated text published to inform the public about matters of public interest. An exception applies where there has been human review or editorial control and a person or organisation assumes editorial responsibility.
Is C2PA enough to comply with the AI Act? Not necessarily. C2PA can provide valuable, verifiable provenance information, but the applicable Article 50 obligation depends on the content, the AI system and the organisation's role. A deepfake, for example, may also require disclosure that is clearly perceivable by the audience.
In practice, ask three questions: what type of content are you generating, could the audience reasonably mistake it for authentic material, and who reviews and takes responsibility for publication? Build provenance into the technical workflow rather than adding it only at the final publishing step, and document meaningful human editorial review where it occurs. Article 50 has applied since 2 August 2026, but its application still depends on the specific circumstances. Neuralex does not provide legal advice: have this reviewed by a lawyer.