Journal

Who has to label an AI ad,
and from when.

Article 50 of the EU AI Act applies from 2 August 2026, with marking obligations extended to 2 December 2026 for systems already on the market. Most of what is written about it confuses two obligations that fall on two different parties.

Tupperware · KettlePRD-001 / 16:9
Tupperware · WhiskPRD-002 / 16:9

Straight off the pass. Both cuts started as a product photograph.

Two separate duties exist and they land on two different parties. The provider of the AI system must mark generated output in a machine-readable way. The deployer, meaning the advertiser publishing the ad, must apply a visible disclosure when the content is a deepfake. A studio producing your ad sits on the provider side of the marking duty; the visible label on the ad you run remains yours.

The two obligations, kept apart

The European Commission's own FAQ separates them explicitly. Providers must ensure generative outputs are marked with machine-readable marks that let the content be detected as generated or manipulated by AI systems. Deployers must apply visible labels to deepfakes and cannot discharge that duty by relying on the machine-readable mark alone.

In practice, for an ecommerce brand running paid social: the machine-readable marking travels with the file we deliver. The visible disclosure on the placement is set by whoever runs the ad account.

The dates

  • 2 August 2026. Article 50 applies.
  • 2 December 2026. End of the grace period that extended marking obligations for systems already on the market before August.

There is no retroactivity on content produced before the obligation applied. That does not mean an old asset can be re-run indefinitely without review.

What counts as a deepfake, and what does not

The Commission sets three cumulative criteria: resemblance to existing persons or things, content that exists in reality or plausibly could, and the capacity to falsely appear to be authentic or truthful. All three must be present.

This is the distinction that matters commercially. A synthetic performer holding your product and speaking a scripted brand claim is realistic content depicting a person who does not exist. A clean product film with no human in frame, generated from your own packshot, does not meet the resemblance-to-a-person criterion in the same way. The two are not the same regulatory object and should not be treated as one.

The editing exemption people forget

The marking obligation does not apply where the system performs an assistive function for standard editing. Colour grading, retiming, upscaling and format conversion are not the same act as generating a person who never existed. Where the line falls on a given asset is a question for your counsel, not for a supplier's blog post.

What the platforms ask on top

Platform policy is separate from the law and changes faster than it. Meta, TikTok and Amazon each expose their own AI-content declaration in the ad creation flow. Two rules survive every version of those policies:

  • Do not present a synthetic performer as a real customer giving a real testimonial. That is an endorsement claim, and in the United States it falls under FTC endorsement guidance regardless of what the AI Act says.
  • Set the platform's own AI declaration flag. It costs nothing and it is the first thing a reviewer checks.

Does disclosure cost you performance?

This is the question every advertiser actually asks. We are not aware of a primary study that isolates the effect of an AI disclosure label on conversion rate for ecommerce video. Anyone quoting a precise percentage is quoting a vendor blog.

What is documented is a perception gap on the advertiser side: IAB research finds a wide gap between how positively ad executives assume younger audiences view AI advertising and how positively those audiences actually report viewing it. Read that as a reason to disclose plainly rather than to hide, not as a performance number.

How we handle it

Every cut we deliver carries the platform AI-disclosure flag pre-set, and we will not script a first-person customer claim that your legal team has not cleared. Scripts are approved in writing before anything is produced. The claim itself stays your responsibility, because only you know what your regulatory position allows.

This page is a summary of published obligations, not legal advice. Verified against the sources below on 2 September 2026.

Sources

Every figure on this page links to a primary source. Where no primary source exists, we say so rather than repeat a number we cannot stand behind.

Need this produced?

Ogamama produces the assets described on this page from product photography you already own, encoded to specification, on a written quote. See today’s service or send one SKU for a free tasting.

Start here

Send one product photo. Taste what comes back.

Tell us what you sell and where you run ads. You get a written plan and a sample cut on your own product, before anything is signed.

  • One SKU, one finished cut, back within two business days. Free.
  • A written production plan: formats, volumes, markets, timeline.
  • No retainer, no lock-in, month to month.

No call required to see the work. The full service log is right here.

By sending this you agree to our privacy notice.