What is a deepfake under the EU AI Act?
The AI Act defines a deepfake in Article 3(60): 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. Two elements do the work — resemblance to something real, and the capacity to mislead. From 2 August 2026, deployers of systems generating deepfakes must visibly disclose that the content has been artificially generated or manipulated — on top of the provider-side machine-readable marking duty under Article 50.
Key boundaries
- Broader than face-swaps: voice clones, staged “event” footage and photoreal renders of real places all qualify;
- Art gets a lighter regime: evidently artistic, creative, satirical or fictional works need disclosure only in a manner that does not spoil the work;
- Consent changes nothing here: agreement of the person depicted does not remove the labelling duty.
The legal dimension
The definition sits in the transparency layer, not the prohibition list: making a deepfake is not banned, hiding its nature is what breaches the Act. The visible label must reach the audience at first exposure; machine-readable marking such as AI watermarking or Content Credentials (C2PA) satisfies the provider’s duty but not the deployer’s. And the label follows the content, not the tool: the same generator’s output may need a label in one context and fall under the artistic exception in another — the general logic of the synthetic content regime.
Turkish context
Deepfakes of real persons engage Turkish personality rights under the Civil Code (TMK) and KVKK — a face or a voice is personal data — even though no AI-specific statute is in force in Türkiye. The November 2025 draft criminal provisions on AI-generated content signal the direction of travel, but remain a draft. For a Turkish startup shipping avatar, dubbing or face tools, the exposure is therefore double: EU labelling duties wherever EU users are reached, and personality-rights and data-protection claims at home.
Do: build the visible label into the output itself — burned in or overlaid — so it survives sharing. Don’t: rely on the depicted person’s consent, or on a watermark alone, to discharge the disclosure duty.
Related guides: Deepfakes and Synthetic Content.
Sources. Regulation (EU) 2024/1689 (AI Act).