Deepfake Labelling Rules Under the EU AI Act

Under the EU AI Act a deepfake is AI-generated or manipulated image, audio or video content that resembles real persons, objects, places or events and would falsely appear authentic. Deployers must disclose that such content is artificially generated or manipulated, clearly and at first exposure.

What qualifies as a deepfake

The test is resemblance plus plausibility: content that looks like real people, places or events and would appear authentic to an ordinary viewer. Fully synthetic scenes of imaginary subjects are treated differently from a synthetic clip of a named politician. Voice cloning counts. So does face swapping, synthetic reenactment, and material edits to real footage that change what a viewer would believe happened.

How the disclosure must be delivered

Disclosure has to be clear, distinguishable and given no later than the first interaction or exposure. That means the viewer learns about it when they see the content, not after clicking through to a description. For artistic, creative, satirical or fictional works the disclosure is lighter — present, but delivered in a way that does not hamper display or enjoyment of the work.

Why a visible label alone falls short

A burned-in corner label is trivially cropped, and it carries no information a machine can check. Article 50 also expects synthetic output to be marked in a machine-readable, detectable form. Platforms, archives and downstream tools need to read the marking automatically. If your only label is pixels, the disclosure disappears the first time the clip is re-cut for a different aspect ratio.

Labelling that survives distribution

Real-world distribution re-encodes, crops, screenshots and re-uploads content. A labelling approach that only works on the original master is a labelling approach that fails in the field. The combination that holds up: embedded credentials where the format supports them, an invisible watermark that survives re-encoding, and an external record that can still be located when both are stripped.

Labelling is not authenticity

A deepfake label states that content was generated or manipulated. It does not state that unlabelled content is genuine, and it cannot certify anyone's intent. Keeping that distinction explicit protects you twice: it keeps the disclosure accurate, and it stops the label being read as a guarantee no technical mechanism can deliver.

Frequently asked questions

Does every AI-edited photo need a deepfake label?

No. The obligation targets content resembling real persons, objects, places or events that would falsely appear authentic. Minor edits that do not change what a viewer would believe are treated differently.

Is a visible watermark enough to comply?

Usually not on its own. A visible label addresses human disclosure; the Act also expects machine-readable marking that software can detect.

Who labels — the tool that generated it or the person publishing it?

Providers must mark generated output; deployers who publish deepfake content must disclose it to recipients. Both duties can apply to the same asset.

Are parody and satire exempt from labelling?

They are not exempt. Disclosure is still required, but delivered so it does not spoil the display or enjoyment of the work.

Check how a label holds up

See what remains verifiable after a file has been re-encoded and re-shared.