What counts as a deepfake under the EU AI Act?

"Deepfake" in the EU AI Act is a defined legal category, and it's both broader and narrower than the word's everyday use.

Broader, because deceptive intent is not part of the test. The duty in Article 50(4) attaches to AI-generated or AI-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. A wholesome AI-generated "photo" for your newsletter qualifies just as much as a malicious impersonation. What matters is that a viewer would take synthetic content for a real recording.

Narrower, because unrealistic content is simply out. Fantastical scenes, obvious stylisation, cartoons, abstract art, in short content nobody would take for a photograph or recording of reality, doesn't meet the "falsely appear authentic" element, however it was made. There is no general duty to label every AI image.

Note also what the definition covers that people miss: AI-manipulated real content, not just fully generated content. A real photo with AI-inserted or AI-removed elements, or a real recording with a voice altered to say new things, resembles real events and would falsely appear authentic, so it's in.

When the duty applies, the disclosure is the visible kind: the harmonised "AI" label at every exposure for images; opening disclaimer plus persistent icon for video; spoken disclaimer for audio, repeated past 30 seconds. For clearly artistic, satirical, or fictional works the duty is attenuated to a non-intrusive disclosure (think credits), but it is not waived, and the carve-out doesn't extend to primarily commercial or informative content.

Text is handled by a different, narrower rule (public-interest publishing, with an editorial-review exemption). The deepfake duty proper is about image, audio, and video.