Everything you need to know about labelling AI-generated content

If you publish AI-generated content, do you have to label it?

Not always. The useful answer starts by separating two obligations that are often treated as if they were the same: technical marking and visible disclosure.

Article 50 of the EU Artificial Intelligence Act has applied since 2 August 2026. It affects both providers of AI tools and the people or organisations using them to publish or deploy content. It does not impose a blanket duty to label every use of AI. It does require a closer look when AI-generated or AI-manipulated content may mislead people about what they are seeing, hearing, or reading. EU AI Act

Marking is not the same as labelling

Technical marking is the provider’s responsibility. It means embedding machine-readable information that can help identify content as AI-generated or AI-manipulated.

Visible labelling, or disclosure, is aimed at people. In the situations covered by the Act, it is the responsibility of the person or organisation deploying the content. Its purpose is to make clear that the audience is encountering synthetic or artificially altered material.

That distinction matters. You cannot assume that a platform’s technical marker resolves your own duty of transparency. Equally, a publisher should not be expected to manually perform a technical task that the law assigns to the provider.

For generative systems already placed on the market before 2 August 2026, the technical marking obligation has a transitional period until 2 December 2026. The visible disclosure obligation in the cases covered by Article 50(4), however, already applies. European Commission FAQ

When images, audio, or video need a label

The best-known obligation concerns deepfakes. But a deepfake is not simply any AI-made asset, or any asset that looks realistic.

The Act refers to AI-generated or AI-manipulated image, audio, or video content that resembles existing people, objects, places, entities, or events and would falsely appear authentic or truthful.

So the practical question is not only whether AI was used. It is this: could someone reasonably interpret this piece as a real record of a person, place, or event?

A synthetic recording in an executive’s voice, an image of an event that never happened, or a video that changes what a person said are obvious examples. In those cases, disclosure must be clear to the people exposed to the content. European Commission transparency guidelines

Artistic, creative, satirical, and fictional works are subject to a more proportionate form of disclosure, one that does not undermine the experience of the work. That is not an exemption from transparency.

Not every AI-assisted text needs a label

This is where much of the confusion lies.

The AI Act does not require a label on every text produced with AI assistance. The obligation applies to AI-generated or AI-manipulated text published to inform the public on matters of public interest, where it has not undergone human review or editorial control and where no natural or legal person holds editorial responsibility.

That distinction matters for media organisations, companies, spokespersons, and communications teams. A text drafted with AI support, genuinely reviewed by a professional, and published under editorial responsibility is not equivalent to an automated public-information piece with no meaningful review.

Review is not changing a few words at the end. It means checking the facts, making editorial decisions, taking responsibility for the claims, and being able to identify who made those decisions.

A label does not replace traceability

Technical markers can be lost when a file is exported, compressed, cropped, or republished. And even when a signal is detected, it does not tell the full story: which source materials were used, what instructions were given, how much human intervention took place, or which version was ultimately approved.

That is why professional teams need more than an icon. They need traceability.

They should be able to answer a few basic questions:

• Which tool and account were used?

• Which source materials were included?

• What was generated or modified by AI?

• Who reviewed and approved the content?

• Where was it published, and was a visible label required?

You do not need to document every minor interaction with an AI tool. You do need a defensible record when content represents third parties, informs the public, is produced for a client, or may create reputational, commercial, or legal consequences.

Five decisions worth making now

First, separate content types in your approval process. Text, images, audio, video, and interactive systems do not create identical risks.

Second, define a criterion for identifying potential deepfakes. Do not leave it to the judgement of the person uploading the final file.

Third, assign editorial responsibility for public-information content. If no one can stand behind the claims, labelling is not the only issue.

Fourth, check which providers offer machine-readable marking and what happens to that information during editing and distribution.

Finally, train the people who create, approve, and publish content. AI literacy is not simply knowing how to use a tool. It is understanding the limits, responsibilities, and contexts that determine whether its use is appropriate.

Transparency should not be a defensive notice added at the end of a workflow. It is a way to preserve trust when the boundary between documentation, simulation, and synthetic content is becoming harder to see.