Deepfake labelling is the obligation to disclose clearly that image, audio or video content has been artificially generated or manipulated so that it falsely appears authentic, together with the technical marking that allows such content to be detected. In the field of Copyright & AI, it belongs to the transparency rules for AI-generated content, and its main legal source in Europe is Article 50 of the EU AI Act, whose transparency obligations apply from 2 August 2026.
The duty is about honesty toward the audience, not about ownership. A label tells viewers that a face, a voice or a scene is synthetic; it does not decide who owns the content, whether it infringes copyright or whether it violates a person’s image rights. Those questions are governed by copyright law, personality rights and data protection, which continue to apply alongside the labeling rules.
What counts as a deepfake under the AI Act
The AI Act defines a deep fake 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. Three elements matter: the content is produced or altered by an AI system, it resembles something real, and it is capable of misleading an ordinary viewer.
The definition is broader than face swaps. A cloned voice in a phone message, a synthetic news photo of a fictional flood in a real city, or a video in which a real politician appears to say words never spoken all fall within it. A clearly stylized cartoon or an obviously fantastical image of a dragon generally does not, because it would not pass as authentic.
Two layers of the obligation
Article 50 splits the work between providers and deployers. Providers build the marking into the system itself, while deployers, meaning the people or organizations that use the system professionally to publish content, must show a visible or audible disclosure to the audience. The table compares both layers.
| Aspect | Provider duty | Deployer duty |
|---|---|---|
| Legal basis | Article 50(2) AI Act | Article 50(4) AI Act |
| Scope | All synthetic audio, image, video and text outputs | Content that constitutes a deepfake |
| Form | Machine-readable marking that is detectable | Clear disclosure perceivable by people |
| Typical techniques | Watermarks, metadata, fingerprints, provenance manifests | On-screen captions, audio notices, labels |
| Technical standard | Effective, interoperable, robust and reliable | Clear and distinguishable at first exposure |
The deployer duty also extends to AI-generated or manipulated text published to inform the public on matters of public interest, unless the text has undergone human review or editorial control and a person holds editorial responsibility for it. Personal, non-professional use falls outside the deployer obligations entirely.
Exceptions and the creative works carve-out
The AI Act recognizes that synthetic media are a legitimate artistic tool. When the content forms part of an evidently artistic, creative, satirical or fictional work, the disclosure is limited to stating the existence of generated content in an appropriate manner that does not hamper the display or enjoyment of the work. A film can therefore carry the notice in its credits rather than across every frame.
Other situations are treated differently, as the following table shows.
| Situation | Disclosure required | Reason |
|---|---|---|
| Political deepfake video in a campaign | Full, prominent label | High risk of deceiving voters |
| Satirical sketch with a synthetic voice | Appropriate, non-intrusive notice | Creative works carve-out |
| Feature film with a de-aged actor | Notice in credits or description | Evidently fictional work |
| Editorially reviewed AI-drafted article | No text disclosure duty | Human editorial responsibility |
| Law enforcement use authorized by law | Exempt | Explicit statutory exception |
| Private family photo edit | Not covered | Non-professional use |
Marking techniques and their reliability
No single technique satisfies the marking duty perfectly, so providers usually combine several methods. Each method trades off robustness, visibility and cost, and the Commission supports a code of practice on marking and labeling AI-generated content to guide providers and deployers toward workable solutions.
| Technique | Strength | Weakness |
|---|---|---|
| Invisible watermark | Survives viewing and light editing | Can be weakened by heavy compression or regeneration |
| Provenance metadata | Rich record of origin and edits | Stripped easily by screenshots or re-uploads |
| Content credentials (C2PA) | Cryptographically signed history | Requires support across tools and platforms |
| Fingerprinting and logging | Matches content against stored outputs | Works only for the provider’s own system |
| Visible label or caption | Immediately understood by people | Cropped or removed with minimal effort |
Detection classifiers that guess whether content is synthetic are useful but not a substitute for marking. Their accuracy drops on new generation models and on compressed media, and false positives can wrongly cast doubt on authentic recordings.
Relationship to copyright and other legal regimes
Deepfake labelling interacts with several neighboring rules but replaces none of them. A label does not legalize content that infringes copyright, for example a deepfake built from a film clip without permission, and it does not cure a violation of a person’s image or voice rights. Equally, the label does not determine the copyright of AI-generated works, which depends on human creative input.
- Training data transparency concerns what went into a model; labeling concerns what comes out.
- Digital Services Act risk duties require very large platforms to mitigate the spread of manipulated media.
- GDPR applies when a deepfake processes the likeness or voice of an identifiable person.
- Other jurisdictions, such as China with its labeling measures and several US states with election deepfake laws, impose comparable but different duties.
Infringements of the Article 50 transparency obligations can lead to fines of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher, with lower caps for small and medium-sized enterprises.
Common mistakes in implementation
Organizations often treat labeling as a single checkbox, yet compliance depends on the whole publishing chain. Typical errors include relying solely on metadata that social platforms strip on upload, placing a disclosure only in a separate terms page, or assuming that a small edit made with an AI tool never counts. A further mistake is labeling authentic content as synthetic out of caution, which erodes trust in genuine material.
In everyday language
A deepfake is a fake photo, recording or video made by AI that looks or sounds real. Deepfake labelling simply means telling people that it is fake. Think of it as the small print on an advertisement that says a scene was dramatized, except that the rules require the notice to be clear at first sight.
There are two parts. The company that makes the AI tool hides an invisible tag in the content so that computers can recognize it later, and the person who publishes the content adds a visible notice for viewers. Jokes, films and art still need a notice, but it can be discreet.
Seen in the real world
A regional broadcaster in the EU produces a documentary about a historical figure and uses a generative model to recreate the person’s voice for several quotations taken from letters. Because the voice resembles a real person and could appear authentic, the segment counts as a deepfake under the AI Act.
The editors place an on-screen caption reading that the voice is AI-generated whenever it is heard, and they repeat the notice in the program description. The voice tool’s provider has already embedded a watermark and signed content credentials in the audio files. When a clip later circulates on social media without the caption, a fact-checking team detects the watermark and confirms that the voice is synthetic, which prevents the clip from being mistaken for an archival recording.