The EU AI Act's label rule for AI ads: what changes from 2 August 2026

Article 50 of the EU AI Act sets transparency duties for two groups. Providers of AI systems that generate content must mark the output in a machine-readable way. Deployers, the people and companies who use those systems, must disclose when image, audio or video content is a "deep fake": content that resembles existing people, objects, places or events and would falsely appear authentic.

For marketing, the second duty is the one that bites. A photoreal model holding your serum, a product shot that never happened, a cloned voice reading the offer: each can fall inside the definition even when no real person is depicted, because the test is whether the content would falsely appear real.

The deployer is whoever uses the AI system under its own authority. Commentators read that as the party deciding how AI is used in the ad. If your agency chooses to generate an AI presenter, the agency is a deployer. If your brand team generates the hero image in-house, the brand is.

In practice both sides should assume they are on the hook and agree in writing who adds the label. The worst outcome is each party assuming the other did it.

Article 50 relaxes the duty for content that is evidently artistic, creative, satirical or fictional, but it does not remove it: the disclosure can be made in a way that does not spoil the work. Advertising is designed to inform a buying decision, so an ad will rarely qualify as evidently fictional just because it is stylised.

There is also an exemption for AI-written text published to inform the public, where a person reviewed it and holds editorial responsibility. It does not cover images, video or audio. Approving a realistic AI image does not remove the duty to label it.

Article 50 itself requires the information to be given in a clear and distinguishable manner, at the latest at the time of first exposure, and the Commission's 2026 guidance and code of practice build on that. That rules out disclosures that appear only on click, sit in a caption below the fold, or flash for a fraction of a second at the end of a video.

Keep the label on the asset itself where you can, in a corner that survives every crop you ship: 1:1, 4:5 and 9:16. Platform AI-label settings help, but they are a delivery mechanism, not proof that your creative met the law.

Inventory which assets used generative AI and how: fully generated, AI-edited, or AI voice. Decide per asset whether a viewer could take it as real. Add the label to every realistic one, in every aspect ratio, and record who decided and when.

That record matters. If a regulator or a client asks why a variant ran without a label, "we checked and it was evidently fictional" is only a defence if you wrote it down before launch.

When a project's markets include the EU, AdVizion flags realistic AI imagery, AI faces presented as customers and synthetic voice without a visible label, pins the finding to the ad, and keeps the decision in the version's audit trail. It is a check, not a legal opinion, so your counsel still owns the final call.