What the transparency rules are about

The EU AI Act approaches advertising from one direction: a viewer should not be misled about whether a person, a voice or a scene is real.

That produces a small number of practical expectations rather than a long list of technical rules. Content generated or substantially modified by AI should be disclosed in a way the viewer actually notices. A synthetic likeness of a real person requires that person’s consent. Disclosure should arrive with the content, not in a footer three clicks away, and it should be in a language the viewer understands.

There is also a direction of travel worth knowing about: machine-readable marking. Provenance metadata — the C2PA-style approach of attaching a signed record to a file — is increasingly treated as the durable way to answer “where did this come from”, because a label can be cropped out of a video while metadata travels with the file.

One clarification that matters commercially: the obligation sits with whoever puts the content into the market. For an ad campaign, that is the merchant advertising the product, not the production partner who made the file. A responsible production partner makes that obligation easy to meet. It cannot discharge it on the merchant’s behalf.

This article describes production practice, not legal advice. For a campaign in a specific market, get the disclosure wording reviewed by your own counsel.

What this means for a product video ad

Most product ads sit in a comfortable middle. They are clearly commercial, they do not impersonate a real person, and they do not fabricate newsworthy events. For those, disclosure is usually a straightforward question of not implying that a presenter is a real customer when they are not.

The risk concentrates in three places.

A synthetic presenter who reads as a real person. If the likeness is modelled on an identifiable individual, you need that individual’s consent, in writing, before production. If the likeness is a constructed persona, the ad should not present it as a customer testimonial — a demonstration and a testimonial are different claims, and only one of them can be produced.

A demonstration that implies proof. Showing a product being used is fine. Showing it producing a result that the product does not reliably produce is a misleading-claims problem before it is an AI disclosure problem. In supplements, cosmetics and health categories this is the claim that gets pulled.

Localised versions. A German cut of an ad made for the US market is still the same ad. If the original needed a disclosure, the localised version needs one too — in German, not in English at the bottom of the frame.

The practical pattern that satisfies all three: decide the disclosure line at the script stage, put it where the viewer is looking, and keep the consent and provenance documents with the campaign file.

Platform labels

Independent of regulation, the major ad platforms run their own AI labelling. Meta and TikTok both provide ways to mark content as AI-generated, and both have moved toward requiring it when realistic imagery or likeness is synthetic. Those labels are set at upload or in the ad creative settings, and they are the merchant’s switch to flip.

Two practical notes. First, a platform label does not replace the disclosure you may owe under local law; they are separate obligations that happen to point the same way. Second, labels are per-asset, so a campaign with twelve variants needs the question asked twelve times — which is an argument for keeping one production file and one disclosure line per campaign rather than improvising per cut.

What HausTrue provides

Four things, by default, for EU-facing work:

  • A flag on AI-assisted production. We tell you which parts of a deliverable were generated or substantially modified, so the disclosure decision is made with facts rather than assumptions.
  • A provenance note on request. A written record of how the deliverable was produced, suitable for attaching to a campaign file or sending to a marketplace or platform that asks.
  • Consent-based asset handling. Merchant assets are not passed to third-party AI services without consent. If a workflow would require it, that is a conversation before production, not after.
  • An AI-free workflow on request. Some categories, and some legal teams, would rather not use generated imagery at all. That option exists at the quote stage rather than as an exception later.

None of this is exotic. It is bookkeeping that most production processes skip because nobody asked for it until an audit did.

A pre-launch checklist for EU campaigns

Run this before the first euro of media spend:

  1. Decide, in writing, whether the ad contains AI-generated or substantially modified content.
  2. If a real person’s likeness is involved, confirm written consent is on file.
  3. Choose the disclosure wording and placement — on the creative where possible, not only in the description.
  4. Set the platform AI label on every variant, not just the hero cut.
  5. Check that localised versions carry the disclosure in the local language.
  6. Re-read the claim language against the product’s actual substantiation, not against the previous campaign.
  7. Keep the provenance note, the consent record and the approved script together, in one folder, for as long as the campaign may run.

That last item is the one that pays for itself. When a marketplace, a platform or a retailer’s compliance team asks how a video was made, the answer is either a document you have or a scramble you do not need.

If you are producing for several EU markets at once, the disclosure and the localisation are the same conversation — one production, several language versions, one consistent disclosure line. That is how HausTrue builds multilingual campaigns, and it is cheaper than retrofitting disclosure into files that have already shipped.