Article 50 and what counts as a deep fake
Article 50 of the AI Act (Regulation (EU) 2024/1689) splits the work in two. Providers of generation tools must mark output in a machine-readable way; the 2026 digital omnibus gave systems already on the market until 2 December 2026 to comply. Deployers, meaning any brand or agency using those tools professionally, must disclose deep fakes: image, audio or video that resembles people, objects, places or events that exist, or plausibly could, and would falsely appear authentic. The Commission's guidance reads that widely, so a photorealistic model who never existed can still need a label. Disclosure must be clear and come at first exposure at the latest. Evidently artistic, fictional or satirical work needs only a disclosure that does not spoil it, and routine assistive edits fall outside the marking duty.
Faces, voices and GDPR
A recognisable face or voice is personal data under GDPR, and a digital twin or voice clone goes further, because it can produce things the person never said or did. Base the work on a written agreement naming the uses, channels, markets, term and fee. If you rely on consent, remember it can be withdrawn at any time, and decide in advance what happens to live assets when it is. National image rights sit on top of GDPR: France, Germany, Spain and Italy each protect a person's likeness under their own law. Where the source recordings are stored also matters. Moving them outside the European Economic Area needs a lawful transfer mechanism, and the recordings are often the most sensitive files in the whole campaign.
Claims, fake reviews and the new green-claims rules
The Unfair Commercial Practices Directive (2005/29/EC) judges the ad, not the tool. A generated before-and-after that shows a result the product cannot deliver is a misleading action however it was made, and fake consumer reviews, including AI-written ones passed off as genuine, sit on the directive's blacklist of practices banned outright. From 27 September 2026, Directive (EU) 2024/825 adds more: generic environmental claims such as 'eco-friendly' or 'climate friendly' without recognised excellent environmental performance to back them, carbon-neutral claims based on offsetting, and sustainability labels not grounded in a certification scheme or set by a public authority. Generated copy produces these phrases unprompted, so block them at the source rather than hoping a reviewer spots them in the fortieth variant.
Platforms, public ad libraries and 24 languages
Under the Digital Services Act (Regulation (EU) 2022/2065), platforms must show who is behind each ad and who paid for it, may not target ads using sensitive data such as health or religion, and may not use profiling to target ads at minors. The largest platforms keep public ad repositories that hold every ad for a year after it last ran, so each generated variant is open to regulators, journalists and competitors. Language is the other constant. The EU works in 24 official languages, and national law adds its own demands: France, for one, requires French in advertising. A campaign that is fluent in English and flat in Czech or Portuguese reads as an afterthought in exactly the markets it was meant to win.
Updated 25 September 2026 · General information, not legal advice. Rules change, so check the current text with your legal team before relying on it.