Disclosure, rights and the EU AI Act before you publish

A beauty brand sent us a cut for sign-off last spring. Everything was in order: product plates, licensed track, a synthetic presenter we had built from scratch. Their legal team asked one question. Where does it say this is AI.

Nowhere. The brief had not mentioned it, and we had not raised it. We added a line to the end card, filed a note of which models made which shots, and the ad ran. It took an hour. It should have taken zero, because it should have been in the brief.

This post is the note I now send with every quote. It is not legal advice, I am a director and not a lawyer, and the rules below are moving. Read it, then check with counsel.

Do AI-generated ads have to be disclosed in the EU?

Broadly, yes, when the content could be mistaken for real. The EU AI Act's transparency rules in Article 50 (the full text is Regulation (EU) 2024/1689 on EUR-Lex) apply from August 2026 and require that AI-generated or manipulated images, audio and video which resemble real people, places or events are disclosed as such, with the people and companies that publish them carrying that duty. On top of the law, the major platforms already require a label on realistic synthetic content in ads.

The practical test I use is the one a viewer would use. Would a reasonable person, watching this on a phone, think they were looking at a camera recording of a real thing. If yes, label it. If it is obviously a stylised render, the law appears to give more room, but the platform may still ask for a label.

Note that the Act's rules on high-risk systems were pushed back in 2026 by the Digital Omnibus package, and the transparency rules in Article 50 were, as I read the coverage, left largely on their original timetable, with some transitional allowances for systems already on the market. Dates have shifted once already, so verify the current ones.

What does Article 50 of the EU AI Act actually require of a brand?

For a brand, the relevant part of Article 50 is the deployer's duty: if you publish AI-generated or manipulated content that looks like real people, objects, places or events, you must disclose that it is artificial. A separate duty falls on the tool makers to mark their outputs in a machine-readable way, which is why watermarks and content credentials are appearing in the files you receive.

The general direction is clear enough to plan around.

• Realistic synthetic people, voices and scenes used in ads need a disclosure the viewer can actually see or hear.
• The disclosure should be clear and distinguishable at the point of first exposure, not buried in a description.
• There is an exception for evidently artistic, creative or satirical work, where the disclosure can be lighter so as not to spoil the piece. Most ads are not that.
• The European Commission published a code of practice on labelling AI-generated content in mid 2026. It is voluntary, but it is the clearest description of what a good label looks like, so it is worth reading.
• Penalties exist and are meaningful. Which tier applies to a transparency breach is a question for counsel.

What do Meta, TikTok and YouTube require?

All three platforms require disclosure when AI-made content is realistic enough to pass as real, and all three now detect and label at least some of it automatically. The common thread is realism rather than the tool: a cartoon does not need a label, a photoreal person speaking to camera does.

Here is my working summary as of this autumn. Policies change often, so treat it as a starting point and read the current help pages.

• Meta. (Meta's approach to labelling AI-generated content.) Content made with Meta's own AI tools is labelled automatically. For content made elsewhere, Meta reads content credentials where present and otherwise expects the advertiser to use the disclosure toggle on photorealistic video or realistic audio. Minor edits such as colour correction are exempt.
• TikTok. (Community Guidelines, integrity and authenticity.) Realistic AI content, including in paid ads, needs the AI-generated label. TikTok has said it may flag undisclosed synthetic content before it reaches viewers, so missing the label is a distribution risk as well as a policy one.
• YouTube. (Disclosing use of altered or synthetic content.) Creators and advertisers must disclose meaningfully altered or synthetic content that looks realistic, through the disclosure setting at upload. Google's stricter mandatory disclosures for ads currently centre on political content, but ordinary commercial ads still fall under the realistic-content rule.

My advice is to disclose on every platform at once rather than tracking three policies.

What about the US: FTC basics

In the US there is no single AI-labelling law for ads, but the FTC's truth-in-advertising rules apply exactly as before: an ad must not deceive, claims must be substantiated, and endorsements must reflect the genuine experience of a real person. A synthetic customer presented as a real one is a fake testimonial, and the FTC has been explicit that AI does not change that.

The 2024 FTC rule on fake reviews and testimonials covers AI-generated ones. So the question to ask of any UGC-style spot is not whether it is AI, but whether it implies a real person had a real experience. I wrote about where that line sits creatively in AI UGC video: when it sells and when it looks cheap.

Several US states also have rules on digital replicas of real people, especially voice, so any spot touching a real likeness is a counsel question first.

Who owns an AI-generated ad?

In the US, purely AI-generated images and footage generally cannot be copyrighted, because the Copyright Office and the courts require human authorship; the human-authored contributions such as the script, the edit, the selection and arrangement of shots, and any photographed plates can be protected. In the EU the picture varies by member state, but the general principle that protection requires human creative choices points the same way.

The case most people cite, Thaler, ended in 2026 when the Supreme Court declined to hear it, leaving the appeals court ruling that a machine cannot be an author. The Copyright Office's 2025 report said much the same: using AI as a tool does not kill protection for the work as a whole, but the raw machine output on its own is not protected.

For a brand this is uncomfortable and simple. The finished, edited, human-directed ad is probably protectable; the individual generated frames probably are not. Your protection against reuse by your own studio or vendor is the contract.

Copyright will not stop anyone reusing your generated frames. Your contract will, so write it down.

Likeness, voice and music: the rights that bite

The rights that actually cause problems on AI jobs are a real person's face, a real person's voice, and music, in that order. Faces and voices are protected by a patchwork of personality, privacy and publicity rights in both the EU and the US, and a recognisable imitation in an ad is the clearest route to a complaint. Music is simpler but routinely forgotten.

On faces: a synthetic presenter built from scratch, with no reference to a real person, is the safe default. If a real face is used as a reference, there must be a written release that covers synthetic reproduction, not just a standard model release from a photo shoot.

On voices: licensed synthetic voices come with terms that specify commercial use, territories and sometimes duration. Read them. A cloned voice of a real actor needs their consent in writing, covering the exact use. We decline briefs that ask for a voice that sounds like a named person.

On music: a library licence for a single online ad typically costs from a few tens of euros to a few hundred, and a composed or buy-out track from a few hundred to a few thousand. AI-generated music carries its own terms and, in some jurisdictions, its own open questions. Whoever licenses it, the licence should name the brand, the use and the term.

References matter too. If a brand sends a competitor's ad or a photograph it does not own and asks for something close, the output may inherit that problem, so we ask where references came from.

What should go in the contract with your studio?

The contract should assign to you whatever rights exist in the deliverables, list the models and tools used with their commercial terms, name who is responsible for disclosure, attach the releases and music licences, and include a provenance log that says which shots were generated, which were photographed and which were composited. If those five things are present you can answer any question a platform, regulator or lawyer asks.

A provenance log sounds bureaucratic. Ours is a spreadsheet with one row per shot, and it takes about twenty minutes to fill in at the end of a project because we keep it as we go.

Of the briefs we received in the past year, only a small share mentioned disclosure or rights at all, which is why this list exists. The paperwork has not caught up with the work.

What I'd tell you on a call

Decide at the brief stage whether the spot will look real, and if it will, plan the label into the creative rather than bolting it on. Keep every face and voice synthetic or released. License the music in your name. Ask for the provenance log and the model list with delivery, and have counsel look at the contract once so you can reuse it.

On a hero spot we budget about half a day for rights paperwork, and it is already inside the ranges in what AI video costs in 2026. The pieces on the front page all shipped with that folder attached.