Your AI ad
has to say so

In fifteen days the rules on synthetic content stop being a discussion and start being law. Article 50 of the EU AI Act applies from 2 August 2026. Most of the coverage has been about what the AI companies have to do. Fine. The part that should worry a marketing department is the other duty, the one that does not sit with the model maker at all. It sits with whoever publishes the work. That is you.

Everyone in this industry has spent eighteen months learning what generative tools can do. Almost nobody has read what happens when you put the output in a paid campaign. Two weeks out, that is a strange place to be standing.

01Two duties, two different owners

Article 50 splits the job. Paragraph 2 is aimed at providers. If a system generates synthetic audio, image, video or text, its outputs have to be marked in a machine readable format and detectable as artificially generated. Watermarks, metadata, cryptographic provenance, whatever works. That is OpenAI's problem, and Google's, and Adobe's.

Paragraph 4 is aimed at deployers. If you deploy a system that generates or manipulates image, audio or video content constituting a deep fake, you have to disclose that the content is artificially generated or manipulated. Same for AI generated text published to inform the public on matters of public interest, unless a human reviewed it and a named person carries editorial responsibility. A deployer is not a lab. A deployer is a brand, an agency, a production company, a social team with a login. The label is yours to put on.

02Narrower than the panic, wider than you would like

Read the definition properly before you go and slap a disclaimer on everything. A deep fake in this Act means image, audio or video which resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. A synthetic cityscape that never existed and does not pretend to is not obviously that. A synthetic version of a real high street, or a real product on a real shelf, or a face that reads as a real customer, is a much harder argument to win.

Two release valves. Where the content forms part of an evidently artistic, creative, satirical or fictional work, the duty is limited to disclosing that the generated content exists, in a way that does not hamper the display or enjoyment of the work. And the provider marking duty does not bite where the system performs an assistive function for standard editing or does not substantially alter the input. Cleaning up a sky is not the same as inventing the whole street.

So no, every retouched frame is not a compliance event. But a photoreal AI hero image of a real looking place, running as paid media across Europe, with nothing on it and nothing in the file, is a bet. Take the bet knowingly or do not take it.

03The number that makes it a board conversation

Breach the Article 50 transparency obligations and Article 99 puts you in the middle tier. Administrative fines up to 15 million euro, or up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start ups the cap is the lower of the two rather than the higher, which is a mercy, but 3% of global turnover is not a line item you explain away in a quarterly review.

Notice what that price is attached to. Not the quality of the work. Not whether the campaign performed. Whether you said what it was.

04The dates, straight

2 August 2026 is when the Article 50 obligations apply. On 10 June 2026 the Commission published the Code of Practice on transparency of AI generated content, drawn up by a multi stakeholder group facilitated by the AI Office, and the Commission and the AI Board have confirmed it as an adequate voluntary route to showing compliance. Signing it is optional. The obligations underneath it are not. Comply another way and you have to prove your way was adequate, to whichever national market surveillance authority takes an interest.

One softener worth knowing. Under the digital omnibus agreed this spring, generative systems already on the market before 2 August get until 2 December 2026 to meet the machine readable marking requirement in paragraph 2, subject to the omnibus being formally adopted. That is relief for the model makers. It is not relief for the deployer duty in paragraph 4, and it is not relief for you. And before anyone in a London office relaxes because this is Brussels law, the Act reaches work whose output is used in the Union. If the campaign runs in Europe, the campaign is in scope.

05What we would actually do about it

Three things, and none of them are legal advice, they are production discipline.

First, log the provenance while you make the work, not afterwards. Every asset in a campaign should have a one line answer to what made this, shot or generated, and who signed it off. Reconstructing that in six months, from four freelancers and a shared drive, is the nightmare. Building it as you go costs nothing.

Second, design the disclosure instead of bolting it on. A label the legal team writes at the last minute looks like an apology. A label you design into the frame from the start reads like confidence. The brands that get this right will make the credit part of the aesthetic, the same way food photography learned to say serving suggestion without anyone flinching.

Third, be honest about when synthetic is the wrong tool. When the whole point of an image is that this is a real place, a real product, a real person who really said that, generating it is not a shortcut, it is a liability with a delivery date. Shoot it. Then you have nothing to declare and a file that proves it.

The era of quietly generating the hero shot and saying nothing has fifteen days left in Europe. That is not a threat to good work. It is a market where owning up is cheap and getting caught is not.