The AI label was in the master. Where did it go in the cut-down?

An AI label is mandatory once your advert contains a deepfake. It is usually right on the master. In the bumper, the square crop and the influencer's repost, it goes missing. Where the label belongs differs from one medium to the next, and radio is the hardest.

Since 2 August 2026, anyone who uses AI to create a deepfake for an advertisement, usually the advertising agency, must disclose under Article 50(4) of the AI Act, visibly or audibly, that the image, voice or video is artificial, at the latest when someone first sees or hears the advertisement, and therefore in every shortened or adapted version too. When something counts as a deepfake has been written about at length in recent months by the European Commission, the Dutch Advertising Code Foundation (Stichting Reclame Code) and VIA Nederland, the Dutch advertising industry association. What happens to the label afterwards has barely been covered.

That is exactly where we see things go wrong. The approved commercial carries the label clearly, after which the media agency turns it into a six-second bumper and someone cuts a square version for Instagram. An influencer gets the material for a post of their own, and three versions later the label is gone. Someone watching that bumper has never seen the master.

Who has to disclose the deepfake in an advertising campaign

The obligation in Article 50(4) of the AI Act rests on the deployer, the party that uses the AI tool under its own authority. In an ordinary campaign that is the agency. According to the guidelines the European Commission published on 20 July 2026, an advertiser that merely commissions the work is not a deployer, as long as it does not decide whether and how AI is used. If the advertiser prescribes AI, or signs off the use of AI image by image, a single campaign may have two deployers.

If the agency has a freelancer operate the tool under the agency’s direction, the agency remains responsible. A realistic generated model can also be a deepfake, even if it resembles nobody in particular. What else falls under it in an advertisement is covered on our page on AI in advertising.

What an AI label looks like in a commercial, on a billboard and on the radio

The law does not prescribe standard wording. At the end of July the Stichting Reclame Code did give examples, such as “Deze stem is met behulp van AI gegenereerd of gemanipuleerd” (this voice was generated or manipulated using AI). According to the Foundation, a vague English line about digital creation will not always say enough. If you would rather use an icon, you can use the AI icons from the voluntary European Code of Practice.

On a still image the label belongs in or directly beside the image, and on video it appears at the start. On a billboard beside the A10 ring road around Amsterdam, legibility is what counts, because nobody reads a line in small-print type from a car.

Radio is the hardest. You hear a cloned voice from the first second. Disclosures in a spot usually come at the end, after the payoff line. By then the listener has heard the voice for twenty seconds and a disclosure comes too late. Put it at the start, or have the voice introduce itself as an AI voice. Where you can foresee that listeners will switch on halfway through the spot and miss that start, the European Commission asks for the disclosure to be repeated later, where possible.

Short video raises the same issue. A label that is on screen for one second in a six-second bumper goes unread. Keep it on screen for as long as the generated person is.

Whether the AI label on Instagram or TikTok is enough

Instagram and TikTok have their own feature for marking content as AI. Useful, but according to the Stichting Reclame Code responsibility stays with the deployer. So check for yourself whether the platform label is visible without anyone having to tap on anything, and whether it sits right next to the advertisement.

The European Commission requires that the audience should not have to take any further action to see the disclosure. A label tucked behind a “more information” menu does not meet that standard. The invisible watermark your AI tool embeds in the file does not count at all.

Where the AI label gets lost along the way: cut-downs, resizes and reposts

A campaign rarely goes out in a single version. The media agency delivers formats for programmatic and social, and every crop can push the label out of frame. A text layer in the bottom right-hand corner often does not survive a square crop.

Reposts are more complicated. An influencer who shares your material without running any AI over it is not a deployer. The European Commission says so expressly about parties that merely disseminate what someone else has made with AI. The obligation therefore stays with the agency that made the deepfake. That agency has no control over what the influencer cuts, but for the influencer’s followers the repost is the first time they see the image. The Commission therefore expects the agency to safeguard the label through arrangements with those who distribute its content.

Running AI campaigns from before 2 August 2026

Material from before 2 August 2026 does not have to be labelled retrospectively. According to the European Commission’s final guidelines, it is the date the material was made that counts, not the date it goes online. An AI campaign from June can therefore keep running without a label, although you may still label it and the Commission encourages you to.

Do record that date, together with the export files and the project folder. If in November you have new images made of the same generated person for a Christmas version, that version falls under the obligation.

Label arrangements in the brief and the delivery specification

You can make those arrangements very concrete. In the delivery specification, state for each asset whether it contains a deepfake, which label goes with it and where it sits. In the brief for the media agency and the influencer, require the label to reappear in every derived version, in a position that survives a crop. How to divide responsibility between agency and advertiser in the contract is explained under the AI clause.

A label does not make an advertisement lawful in itself. If a generated customer tells viewers how pleased they are with your product, the advertisement may be misleading (Article 8 of the Dutch Advertising Code and Article 6:193c of the Dutch Civil Code). Label or no label.

If you are working on a campaign with an AI model or an AI voice, we will go through your delivery list and your arrangements with the media agency and influencers with you before the first version goes live. Get in touch or call 020 675 88 21.