EU AI Act Deepfake Labeling

Short answer

Under Article 50, deployers of AI systems that generate or manipulate image, audio or video content constituting a deepfake must disclose that the content was artificially generated or manipulated. That disclosure is framed as clear information for people who encounter the content, subject to exceptions described in the Regulation and Commission guidance — including certain artistic, creative, satirical or analogous work, and authorised law-enforcement uses. It is a different duty from the provider obligation to apply machine-readable marks to synthetic outputs. A file can carry a machine-readable mark and still raise separate deployer labelling questions when published as a deepfake; conversely, a visible label does not automatically fulfil provider marking rules.

Deployer disclosure versus provider marks

Providers place generative systems on the market and, under Article 50(2), must pursue machine-readable marking and detectability of synthetic audio, image, video or text as far as technically feasible. Deployers use systems under their authority. For deepfakes, the transparency burden highlighted in Commission materials is the deployer's duty to disclose artificial generation or manipulation so people can recognise it. Mixing “watermark present” with “labelled for viewers” collapses two layers the Act keeps distinct.

Image, video and audio context

Deepfake rules in Article 50 are tied to image, audio or video content that constitutes a deepfake as defined in the Regulation. Text published on matters of public interest has related but separate deployer disclosure rules. This page does not redefine “deepfake”; official definitions and guidelines control.

Artistic and other exceptions

Official materials describe limited exceptions — for example where disclosure would be incompatible with the exercise of fundamental rights and freedoms, including freedom of expression and the arts, while still requiring appropriate information about the existence of generated or manipulated content in a manner that does not hinder display or enjoyment of the work, and law-enforcement authorisations where applicable. Exact wording and application belong to the legal text and competent authorities. This site restates the existence of those categories; it does not invent additional carve-outs.

Limitations

• No political advocacy or recommended policy position.
• No penalty interpretation or case-outcome prediction.
• AI Watermark Center tools do not certify deepfake labelling compliance.

Related reading

• Article 50 page
• Machine-readable marking
• Code of Practice