A deepfake is AI-generated or manipulated image, audio or video content that closely resembles existing people, objects or events. What matters is its capacity to deceive — not the technology behind it.
The term comes from "deep learning" and "fake". Colloquially it usually means doctored celebrity videos. Legally it is now narrower: the EU AI Act defines deepfakes in Article 3(60) and attaches a disclosure duty to them in Article 50(4).
Deepfake in detail #
The distinction that most often goes wrong in practice: not every AI image is a deepfake. A wholly invented motif with no real-world referent is AI-generated content under Article 50(2) — it must be labelled, but it is not a deepfake. Content becomes a deepfake only by rendering something that exists in a deceptively realistic way.
In practice: the invented cyberpunk image in an article header is not one. A video in which a real executive says things they never said is one — even if both came out of the same tool.
Ask yourself: could someone take this for a recording of reality? If yes, treat it as a deepfake and label it accordingly — regardless of how you meant it.
What has applied since 2 August 2026 #
Deepfakes are not banned. The EU AI Act asks for transparency instead of prohibition: whoever publishes them must disclose that the content was artificially generated or manipulated. No specific wording is prescribed — it only has to be clearly and distinguishably recognisable.
Personality rights, copyright and criminal law apply regardless. Labelling does not make unlawful content lawful — it only satisfies the transparency duty.
Common questions #
Are deepfakes banned?
No. The EU AI Act requires disclosure, not prohibition. The duty sits in Article 50(4) and has applied since 2 August 2026. Criminal and personality-rights limits exist independently.
Is every AI image a deepfake?
No. A purely fictional motif is AI-generated content under Article 50(2) — labelled, but not a deepfake. The deceptively realistic depiction of something that exists is what makes the difference.
How do I label a deepfake correctly?
Visible in the image plus caption; on screen at the start of a video and repeated if long; spoken at the start of audio. In German-speaking markets "KI" reads more clearly than "AI" and is therefore the safer choice.
Does this apply to older content?
Content finished before 2 August 2026 does not have to be labelled retroactively per the Commission guidelines (paragraph 154). The moment you re-edit it, a new version is created — and that one falls under the duty.
Sources #
- Regulation (EU) 2024/1689 (AI Act) — Art. 3(60) definition, Art. 50(4) disclosure
- European Commission, guidelines on Art. 50 AI Act, 20 July 2026 — section 8.4, paragraph 154 (entry into application)
- Wettbewerbszentrale — guide to AI labelling (February 2026)
Not sure whether your material counts?
Ask me — or ask Ada. She is my AI assistant, available around the clock, and will propose a slot with me if you need one. Free, no form.