The EU AI Act is live. Does a little AI in your ad require a label? What Enterprise Marketing Teams should know.
Executive Summary
- Article 50 of the AI Act has applied since 2 August 2026. Infringement carries fines up to 15 million euros or 3% of worldwide annual turnover, whichever is higher.
- The disclosure duty turns on whether the AI-generated element changes what a viewer believes they are looking at, however small a share of the asset it occupies.
- A real product photographed against an AI-generated background is not a deep fake. Ask a generative model to fix the lighting and it may quietly change the product, which makes it one.
- Deployer status turns on who controls how AI gets used in production, which makes it a question for your agency contracts.
- Most marketing copy sits outside the text rule entirely. Claims about health, consumer safety or sustainability do not, and a second regime lands on those on 27 September.
The AI Act requires you to tell people when content was generated or altered by AI. That duty started binding marketing teams on 2 August 2026, when Article 50 began to apply. In this article, you will read about which content it covers, and what the label has to look like. Every claim below is quoted from the primary text and linked to where it sits.
Are you a deployer under the EU AI Act?
Article 50 places the labelling duty on the deployer, defined in Article 3(4) as a body "using an AI system under its authority." The provider is whoever built the model you are calling: OpenAI, Google, Midjourney, ElevenLabs. Everyone downstream of them looks like a deployer, and the Commission carves out one case:
"By contrast, a company that merely commissions an advertising agency to produce an advertisement, without taking decisions and exercising control over whether and how the advertising agency uses AI in the production process, is not a deployer." European Commission Guidelines (2026)
An AI usage policy is control. Vetoing a tool on security grounds is control. Reviewing generated assets before they ship is control. So for most enterprises running campaigns the duty is yours.
Article 50(4), first subparagraph, is the duty. Deployers of a system that generates or manipulates image, audio or video content "constituting a deep fake, shall disclose that the content has been artificially generated or manipulated."
Article 50(5) governs manner and timing. The disclosure goes to the people concerned "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure," and "shall conform to the applicable accessibility requirements."
Article 50(2) belongs to your model vendor. It requires them to mark outputs in a machine-readable format, and it matters to you only because it does not cover you. Pointing at a vendor watermark as proof of your compliance cites someone else's duty, and the Commission closes that door explicitly: deployers "cannot rely on the machine-readable marking embedded in the content by the provider under Article 50(2)," because those markings "are not immediately clear and distinguishable" to the people looking at the content.
What counts as a deep fake in advertising
Article 3(60) defines a deep fake as AI-generated or manipulated image, audio or video content "that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful."
The Commission judges the finished piece as a whole: how closely it resembles the real thing, what it appears to be saying, where it will be shown, and who is likely to see it. Your intent plays no part, since the assessment "is objective and does not require the intention of the deployer to deceive or mislead."
What decides the answer is which layer of the frame the AI touched.
What decides the answer is which layer of the frame the AI touched.
So an AI-generated sphinx over the Eiffel Tower is not a deep fake, and neither are animated mice arguing about cheese in an ad for a cheese manufacturer.
Both cases are clearly imaginative. A synthetic avatar of your own CEO is one. And so is this:
"An AI-generated image of a product in advertisement or packaging that can affect the audience's perception and mislead as to the actual product appearance, characteristics or use (e.g. making the product appear not identical to the real product, more appealing or with improved quality than in real life)." European Commission Guidelines (2026)
There is no minimum. Three seconds of a synthetic presenter inside an otherwise real sixty-second film triggers the duty, exactly as a fully synthetic film would. The Code of Practice takes this for granted, describing an AI segment that "does not necessarily span the entire duration of the video."
Does AI colour correction count as a deep fake?
The Commission's safe list describes results. Most AI tools do not work in results, and that is where this gets difficult.
Ask a generative model such as Gemini's Nano Banana to colour correct a photograph of a real car and it does not move a slider over your pixels. It produces a new image. Usually the result is a colour corrected car. Sometimes you also get a slightly different wheel, a softened panel gap, a badge that reads not quite right, or a roof rail that was never on the original vehicle.
Regeneration is not itself the trigger. The Commission's own safe list includes converting a black and white image to colour.
The trigger is the result no longer matching the real product. At that point you have an image "making the product appear not identical to the real product", and the duty applies. Your prompt offers no protection, because the assessment "is objective and does not require the intention of the deployer to deceive or mislead." You asked for a colour correction. What gets assessed is what came back.
Which converts a settled legal question into an evidentiary one: can you show that your colour correction stayed a colour correction? An exposure slider is bounded by arithmetic and cannot invent a wheel. A regeneration is bounded by nothing, and the drift is usually invisible until someone puts the two files side by side.
Does AI-generated marketing copy need a label?
Article 50(4), second subparagraph, covers text "published with the purpose of informing the public on matters of public interest." The Guidelines put company marketing copy outside that phrase: "AI-manipulated text that is part of a company's advertisement or product descriptions" is explicitly out of scope, as are internal communications, because organisation-internal texts on corporate networks are not "published" for these purposes.
Your product copy, your campaign headlines and your internal newsletter are not caught by this rule.
There is one exception, and it catches enterprise marketing. Claims about health, consumer safety or sustainability stay in scope.
Take one line of packaging copy: "made with 80% recycled plastic." AI-translate it into eleven markets and you have eleven AI-manipulated sustainability claims, all of them in scope. A second law reaches the same eleven sentences on 27 September 2026, when the Empowering Consumers for the Green Transition Directive starts to apply.
Where the AI label goes, and what it can look like
The EU publishes three icons, each in black, white, and both at 50% transparency.
- Basic icon: "When AI was involved in the creation of deepfake content (image, audio, video) or published text."
- Fully AI-Generated: "When the entire deepfake content (image, audio, video) or text is fully generated by AI."
- Partially AI-Modified: "When pre-existing, human-made content was partially modified with AI."
Their use is optional, and the Commission is direct about it: "The use of these EU icons is optional, but the labelling requirements under Article 50 AI Act are not."
Anyone can use these icons, free and without crediting the Commission. You can also design your own, provided it meets the same specifications, which the Commission sets out in its Code of Practice.
They are narrower than they sound. The mark carries the capitalised acronym AI in English as its main visual element, with letters of equal height and proportions preserved when resized. The acronym may appear in a national language only where English "is incompatible with applicable national laws on the use of languages in commercial or administrative matters." France is the clearest exception, since the Toubon law has required French in commercial communications aimed at the French public since 1994.
Everything else is open, and the Code says so directly: the label "may be expressed in different styles (e.g., contrast ratio, colour, or typography), as long as it remains clear, accessible, and distinguishable." A disclosure mark does not have to look like a government badge stapled to your creative. It can be built to belong to the brand it sits on.
Placement follows the format, and the Code sets it out per modality.
Read together, those two provisions leave real design room. The mark has to carry the AI acronym, stay legible, and keep clear of whatever else is in frame. Colour, contrast, typography and size are yours, which means the disclosure can be drawn in your own brand system rather than pasted on as a foreign object, provided it stays recognisable as a disclosure.
One placement rule is really an engineering requirement. The label must keep "sufficient spacing to other overlay elements, disclaimers, and sound/visual on-screen/on-display elements" and "remain visible against any background." In plain terms, it cannot sit on top of anything else, and it has to stay readable whatever is behind it.
So the label cannot live at one fixed corner. In one market that corner holds a broadcaster logo. In another it holds a subtitle track. In a third, a retailer's promotional banner. The label has to move.
How to prove which layer of an asset was AI-generated
Every determination above resolves to the same two facts about a finished asset.
The first is which layer came from where: whether the generated element was the background or the product, whether the presenter was filmed or synthesised, whether the sustainability sentence was written by a person or by AI.
The second is when the human review happened. A text claim that relies on human review to avoid a label loses that protection if AI edits the copy again afterwards, so the order of the steps decides the outcome.
A flattened export answers neither. Open a finished MP4 or a delivered JPEG and there is no way to recover which layer was generated, when it appeared, or what happened to the copy after somebody approved it. The information existed during assembly and stopped existing at render. A vendor watermark does not close the gap, because platform re-encoding, cropping and colour grading routinely might strip it, and it was never designed to say which layer of a composite it refers to.
This is the same problem enterprises already solved for a different reason. A content supply chain that produces thousands of correct variants of one product already has to know which module built which layer, for which market, in which format. That record is what makes personalization and localization work at volume. It is the identical record that answers the compliance question, because both are asking which component came from where.
We build that infrastructure, and the provenance comes out of it rather than being added later. One campaign ran 15,000 assets across 38 products, cutting a 7,500-hour manual process to under 94 automated hours. Another rollout reached more than 20 markets. Every rendered variant carries the record of how it was assembled, because the system had to know that anyway to render it correctly. The same property that produces the speed produces the audit trail.
Little of what it assembles is AI-generated and needs a label. Most is fetched: the sustainability claim pulled from the PIM as approved text, already translated and signed off by a named team; the product shot pulled from the DAM as the photograph that was actually taken. Generation runs where it earns its place, and gets labelled accordingly. Everywhere else the system retrieves the approved asset and records which one it retrieved.
"This sentence is claims library entry 4471, human-approved, unchanged since..." answers a regulator. "A model produced this and we believed it stayed faithful" does not.
The system is a Content Supply Chain and is called VARYCON, and we are building Article 50 into it now: which elements need a label, which mark to use, and where it goes for each market and format. It needs no new data. The record that answers a compliance question is the one the system already uses to build the variant.
If your brand is working out which assets need a label, and you have to produce them across markets and formats at volume, we are happy to talk it through and show you the options for building it in a way that holds up.
Related Questions
Does the EU AI Act require me to label all AI-generated marketing content?
No. For images, audio and video, the duty applies only where the content constitutes a deep fake under Article 3(60), meaning it resembles real persons, objects, places or events and would falsely appear authentic. AI-generated backgrounds, colour correction and aesthetic compositions in product advertising generally fall outside it. An AI product image that misrepresents the actual product does not.
Does AI colour correction or lighting adjustment count as a deep fake?
Not on its own. The Commission treats colour correction, lighting adjustments and aesthetic background replacement as having only a minor impact on perceived authenticity, and the test measures the result rather than the technique, so regenerating pixels is not itself the trigger. The exposure comes from drift: a generative model asked to colour correct a product photograph may also alter the product, and an image that makes a product "appear not identical to the real product" is a deep fake regardless of what the prompt asked for.
Is my brand or my agency the deployer under Article 50?
Whichever one exercises authority and control over how the AI system is used in production. The Commission's Guidelines state that a company which merely commissions an agency, without taking decisions about whether and how that agency uses AI, is not a deployer. Both parties can be deployers simultaneously where both exercise control, so the allocation belongs in the contract.
Does AI-generated advertising copy need a disclosure under the AI Act?
Generally no. The text duty in Article 50(4) covers only text published to inform the public on matters of public interest, and the Commission explicitly places company advertisements and product descriptions outside that scope. The exception is copy making claims about health, consumer safety or sustainability, which returns to scope, as does investor-facing corporate reporting.
What is the best way to make AI-drafted text compliant at scale?
Stop regenerating the sentences that carry the risk.
Most of your copy is out of scope, so the exposure sits in a small set of claim types: sustainability, health, consumer safety, and investor material. The catch is that those claims do not stay in one file. A single sustainability line on a master is inherited by every localised version of it, so one claim can become hundreds of in-scope assets.
Treat those claims as fixed content. Hold them in an approved library, already translated and already signed off, and have the system place them rather than generate them. Nothing is rewritten after approval, so the exemption survives across every variant.
Where you genuinely need generation, the human review has to come after the last AI step, not before it. At volume that only works if the system can identify which variants contain a regulated claim and route those to a reviewer who understands it. Nobody finds them by opening files and reading.
Citations
1. Regulation (EU) 2024/1689, Article 50, European Commission AI Act Service Desk (2024): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
2. Regulation (EU) 2024/1689, Article 3, European Commission AI Act Service Desk (2024): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-3
3. European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act, C(2026) 5054 final (2026): https://ai-act-service-desk.ec.europa.eu/sites/default/files/2026-07/guidelines_on_the_implementation_of_the_transparency_obligations_for_certain_ai_systems_under_article_50_of_the_ai_act_bzptwqhk0ikg1dtlddap41psfy_131215.pdf
4. European Commission, Code of Practice on Transparency of AI-Generated Content (2026): https://ec.europa.eu/newsroom/dae/redirection/document/129555
5. European Commission, EU icons for labelling AI-generated content (2026): https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content
6. Regulation (EU) 2024/1689, Article 99, European Commission AI Act Service Desk (2024): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99
