Labelling AI-generated ads in Europe: what Article 50 of the AI Act and France's ARPP require

In one sentence: Since 2 August 2026, an ad shown in the EU whose image, voice or video was generated or manipulated by AI and could pass for real must say so; the European Commission's guidelines and the ARPP's practical sheet of 3 August 2026 give a case-by-case method to decide.

This guide, part of the Guides section, is for brands, agencies, in-house marketing teams and freelancers who produce campaigns with generative AI and show them to audiences in the European Union. It rests on four official texts: Article 50 of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689), the Commission guidelines on Article 50 dated 20 July 2026, the Commission's code of practice and EU icons, and a practical sheet (in French) from the ARPP (Autorité de régulation professionnelle de la publicité), France's advertising self-regulatory organisation. The ARPP sheet is French, but it applies a method designed at European level, so its grid is useful well beyond France.

Who is covered, inside and outside the EU

Article 50(4) of the AI Act places the duty on "deployers". The regulation defines a deployer as a person or organisation "using an AI system under its authority", except for purely personal, non-professional use. The company that develops and markets the AI tool is a "provider", with separate duties such as machine-readable marking.

Location does not get you out of it. According to the Commission guidelines, a deployer established outside the Union is covered when it foresees that the output will be used in the EU, by directing or authorising distribution there, which includes posting the content on the globally accessible internet. The guidelines give the case of a company in a third country that uses AI to create a deepfake of a celebrity for an ad displayed in the Union: it is a deployer within scope. A third-country deployer is not bound when content reaches EU audiences through channels it could not foresee or control. The ARPP sheet makes the same point for advertising: the deployer can be an advertiser, an agency or another operator, inside or outside the EU, as soon as the content is distributed in the Union.

Between brand and agency, the test is decision-making power. The guidelines on Article 50 say employees acting on the instructions of a company (designers, animators, content creators) are not separate deployers: the company is, and it remains the deployer when it brings in contractors or freelancers under its control. Conversely, a company that "merely commissions an advertising agency to produce an advertisement", without deciding whether and how the agency uses AI, is not a deployer. Write it into the agency contract: who decides on AI use, who checks the final cut, who places the label. The same deployer logic structures the AI Act's rules for employers, covered in our guide to AI CV screening under the CNIL and the AI Act.

What Article 50 requires, and since when

The AI Act defines a deepfake (Article 3(60)) 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". Deployers who generate or manipulate such content "shall disclose that the content has been artificially generated or manipulated". For evidently artistic, creative, satirical or fictional works, the duty is lighter: disclosure "in an appropriate manner that does not hamper the display or enjoyment of the work". A second subparagraph covers AI-generated text "published with the purpose of informing the public on matters of public interest", unless it has undergone human review or editorial control and someone holds editorial responsibility. Article 50(5) adds that the information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure", and must meet accessibility requirements.

The dates are worth getting right, because a July 2026 amendment moved several AI Act deadlines. Article 113 of the AI Act sets 2 August 2026 as the general date of application, and the Commission guidelines confirm that Article 50 applies from that date. Regulation (EU) 2026/1744, the "Digital Omnibus on AI" published in the Official Journal on 24 July 2026, did not move the labelling duty for deployers. It gives only providers of generative AI systems already on the market before 2 August 2026 until 2 December 2026 to comply with the machine-readable marking of Article 50(2). For an advertiser, that grace period changes nothing: the guidelines add that deployers cannot rely on the machine-readable marking embedded by the provider, because it is not "immediately clear and distinguishable" to people. The full sequence of dates is set out in our EU AI Act timeline after the 2026 delay.

Two practical points follow from the guidelines. Content generated or manipulated before 2 August 2026 does not have to be labelled retroactively, and nobody is expected to audit old archives or reprint packaging. Penalties for breaching Article 50 reach 15 million euros or 3% of total worldwide annual turnover, whichever is higher, and for SMEs and start-ups whichever is lower (guidelines, section 8.3, Article 99(4) of the regulation). Enforcement belongs to the market surveillance authorities designated by each Member State.

The French layer: what the ARPP is, and is not

The ARPP is not a public authority. Its practical sheet (in French) says so itself: neither the ARPP nor the JDP (Jury de déontologie publicitaire, the advertising ethics jury linked to the ARPP that hears complaints about ads) is "the competent authority" for enforcing the AI Act, and at the date of the sheet the French authorities in charge were "not yet formally designated" (our translation). Their designation was planned by a bill adapting French law to EU legislation (projet de loi portant diverses dispositions d'adaptation au droit de l'Union européenne, or DDADUE), then before Parliament. Until then, the sheet serves as a reference for the advice the ARPP gives to professionals, and for its future opinions once national enforcement rules are in force.

What makes the sheet useful outside France is its lineage. It states that it follows the Commission guidelines and the work of the European Advertising Standards Alliance (EASA), the European network of advertising self-regulatory bodies, which developed the three-factor approach, and that it adapts this approach to the French context. It also recalls that existing law, the ARPP's own code of recommendations and the ICC Code on advertising and marketing communications apply whether or not the content is AI-made.

When an ad needs a label: the three-factor grid

The ARPP sheet starts from a clear principle: "using artificial intelligence does not, in itself, create an obligation to inform the public" (our translation). AI used in preparatory stages (ideas, script, storyboard, translation, optimisation, technical retouching) does not, in principle, call for any specific notice. A label is needed when the ad as broadcast is generated or manipulated by AI and is likely to be perceived as authentic or truthful. Only the content actually distributed counts.

To decide, the sheet weighs three factors together:

1. Level of AI intervention: low (minor technical changes), intermediate (substantial manipulation) or high (complete or near-complete generation).
2. Realism: a stylised or unreal universe is low-stakes; a realistic depiction of people, places or events is high-stakes.
3. Effect on the consumer: does the AI element change how the audience perceives or understands the message or the product?

"None of the three principles is sufficient on its own", the ARPP adds (our translation). The test is whether one or more AI elements significantly affect how the consumer perceives the ad. The guidelines also ask deployers to think about the reasonably foreseeable audience, including children and older people, who may be more easily misled.

On creative work, the two texts differ in tone. The ARPP writes that "no exception is provided for advertising" (our translation) in the lighter regime for artistic and fictional works. The Commission guidelines are slightly more open: advertisements "might be regarded as evidently creative or fictional works in certain, specific situations, but not in others", and when content mixes creative and informative character, "the informative character should always prevail". In practice, treat a realistic ad as a standard case.

Common advertising cases, side by side

The case lists of the ARPP sheet and the Commission guidelines overlap on the situations agencies meet most often; where the Commission gives no matching case, the table says so.

The ARPP adds two cases that trip up campaigns: an AI-improved sound for a children's toy, and a realistic reconstruction of an event presented as if it had been filmed. Both call for a label. The Commission lists a teleshopping-style video with simulated consumers, and a realistic synthetic influencer testing a sponsored product, among deepfakes that do not count as creative works.

Advertising copy and other text

Text follows a different rule. The Commission guidelines place AI-manipulated ad copy and product descriptions outside Article 50, unless they include claims related, for instance, to health, consumer safety or sustainability. Corporate communication is another matter: AI-manipulated reports with investor information on a listed company's website are given as text informing the public on matters of public interest.

For such text, the exemption requires real human review, and the guidelines are strict about it. Spell-checking, the mere existence of an editorial policy, automated checks or a cursory sign-off "cannot fulfil the conditions", and any substantive AI intervention after sign-off makes the exemption void. The person or function holding editorial responsibility should be identifiable, for instance in a website's legal information.

How to word and place the label

The ARPP sheet asks for a label that is visible or audible, clear and distinct from other information, readable for long enough, accessible and suited to the medium, and that stays attached to the content when shared. For video, a short text notice appears from the start of the message; for audio-only ads, a short spoken notice; an intelligible icon also works. Its suggested wordings translate as "AI-generated image", "AI-generated voice", "AI-generated character", "AI-generated video" and "AI-generated advert" (our translation), offered only as guidance: no standard wording or icon is imposed by the AI Act or the Commission guidelines.

The guidelines add two placement rules. A notice buried in terms of use or behind menus does not count as clear and distinguishable. And if viewers may reasonably not see a video from the beginning, a label at the start alone is not enough and should be repeated later where possible.

For a ready-made visual, the Commission offers the EU icons for AI-generated content, free to use without attribution, in black, white and two semi-transparent versions. There is a basic icon, a "Fully AI-Generated" variant and a "Partially AI-Modified" one; user testing found they work better with a short text label. The icon should be embedded in the content, visible at first exposure and still visible when the content is reshared or downloaded. Using the icons is optional, and they do not establish compliance by themselves.

The icons belong to the Code of Practice on Transparency of AI-generated Content, published on 10 June 2026, whose second section sets labelling rules for deployers. The code is voluntary, but the Commission and the AI Board have confirmed it is an adequate tool to demonstrate compliance, and the Commission reports about 190 signatories by the end of July 2026. According to the guidelines, a non-signatory must show its measures are adequate by other means, for instance with a gap analysis against the code, and can expect more requests for information from authorities.

What a label does not fix

A label is not a licence. "Misleading communication remains misleading, even when labelled as using artificial intelligence", the ARPP sheet warns (our translation): AI that exaggerates a product's performance falls under the rules on misleading commercial practices. The Commission guidelines likewise say Article 50 does not make lawful a deepfake that is unlawful under other rules, such as misleading advertising law. When a real person's face or voice is reproduced, you still need their prior consent, a legal basis under the GDPR and respect for personality rights; when protected works or trade marks appear, intellectual property law applies too.

Using a chatbot to apply the grid

A chatbot can help you run a campaign through the grid, provided you give it the criteria and keep the decision. These tools can produce plausible but wrong answers, so the conclusion belongs to the person responsible for publication. The template below is a starting point to adapt; Le Latent has not run it on a real campaign.

I am preparing an advertising campaign for distribution in the European Union. Here are the elements generated or modified by AI in the final version the public will see: [list: main visual, voice-over, extras, background, headline copy].
For each element, apply these three criteria and answer in a table:
1. Level of AI intervention: low (technical retouching), intermediate (substantial manipulation) or high (complete generation).
2. Realism: does it depict people, objects, places or events that exist or could plausibly exist?
3. Possible effect on the consumer: could it make the message or the product appear authentic when it is not?
End each row with "label to consider", "label probably not needed" or "to be decided", and never present your answer as legal advice.

A second template, also to adapt, drafts the notice for each format.

Suggest a short notice stating that [element: the image / the voice / the character] was generated by AI, in versions for: a poster, a 15-second video (on-screen text at the start, repeated at the end), an audio spot, and a social media post. Language: [language of the target market]. Neutral tone, plain words, no promotional wording.

To follow how national authorities are designated and how Article 50 is enforced, read the Society section, and for other AI Act changes affecting your work, browse more practical guides for professionals.