AI Act and advertising: what advertisers actually have to do

Since 2 August 2026, labelling AI-generated advertising content has been an enforceable obligation. Which types of AI-assisted production are in scope? What obligations do advertisers face? Here is what you need to know.

“Image generated with the assistance of AI.” You have been seeing this short line more and more often since August 2026: at the bottom of a billboard in a metro corridor, overlaid during the first seconds of a TV spot, under a sponsored post, on the back of a pack…

It makes Article 50 of the AI Act visible to everyone. The article has applied since 2 August 2026, and breaches can be fined up to €15 million or 3% of total worldwide annual turnover (for SMEs, whichever amount is lower applies). Little wonder the notice is spreading across advertising visuals.

But before adding it, advertisers have had to answer three questions:

• Does the creation of the visual or video fall within the scope of the text?
• Who decides to label it, and who carries it out?
• Who is accountable?

In this article

• What has applied since August 2026
• Good news: AI targeting is not banned
• Labelling: a far broader definition than deepfakes
• Adding the “made with AI” notice: where and when?
• Advertiser or agency: who labels?
• Four off-the-radar issues not to underestimate
• The AI Act does not stand alone

What has applied since August 2026

First, a look back at the text itself, as there has been some confusion about what applies now and what has been postponed. The Digital Omnibus on AI, in force since 27 July 2026, did push back certain obligations*: those attached to so-called “high-risk” systems. This label refers to use cases that can harm people’s health, safety or fundamental rights.

The transparency requirements that apply to advertising communications, among others, are very much in effect: Article 50 of the AI Act applies, and authorities have been entitled to penalise non-compliance since August 2026.

In France, the supervisory authorities have not yet been designated: the DDADUE bill**, passed by the Senate in February 2026, is awaiting the National Assembly. Under the planned framework, Arcom (the French audiovisual and digital communications regulator) would oversee the transparency of AI-generated content, and the CNIL prohibited practices. Neither the ARPP (France’s advertising self-regulatory body) nor the Jury de Déontologie Publicitaire would play that role. The obligation nevertheless applies now, even if the tools for administrative enforcement are still being put in place. For a few more months, the risk lies more in reputation and civil litigation than in fines. The urgency to comply, however, is undiminished.

Good news: AI targeting is not banned

Let’s start with some good news. A lookalike audience built with machine learning, a retargeting sequence optimised by an algorithm, a recommendation engine: none of these is prohibited.

The Commission says so explicitly in its guidelines on prohibited practices: tailoring an advertisement to a person’s preferences is lawful persuasion. The first part of Article 5 of the AI Act only prohibits an advertising practice if four conditions are met at the same time, so the absence of just one is enough to rule it out. They are:

• an AI system within the meaning of the Regulation, which excludes segmentation based on manually written rules;
• a subliminal, manipulative or deceptive technique: “subliminal” means a stimulus below the threshold of conscious perception, such as a message flashed too briefly to be seen or an inaudible frequency, while “manipulative” refers to bypassing a person’s capacity for rational decision-making, not to persuasion;
• a material distortion of behaviour, i.e. a decision the person would not otherwise have taken;
• significant harm (physical, psychological, financial or economic) linked to the practice by a plausible causal relationship.

These criteria (especially the last three) put ordinary commercial targeting out of reach: the Commission considers that adapting an ad to a person’s preferences is lawful persuasion, and that legitimate advertising practices are not inherently manipulative. The harm criterion then acts as a second safeguard, and it explains why the prohibited cases focus on situations of vulnerability. There are explicit prohibitions, but they concern specific cases:

• targeting over-indebted people with short-term credit offers;
• messages built on personal data to exploit an individual’s vulnerability;
• sound or visual environments designed to heighten anxiety.

Note that intent to cause harm is not required, and that the prohibition covers systems that have learned manipulative techniques during training. In practice, a generative testing setup that optimises its own creatives and drifts towards anxiety-inducing triggers makes the advertiser liable, even if nobody intended it.

Labelling: a far broader definition than deepfakes

The metro poster features no famous face, yet it carries a notice. Excessive caution? Not really. The reason lies in how the European text defines a “deep fake”, which has little to do with the everyday meaning of the word.

It is defined as image, audio or video content, generated or manipulated by AI, that appreciably resembles existing or plausible persons, objects, places, entities or events, and is likely to falsely appear authentic.

The notion of objects explicitly covers consumer goods. The Commission’s guidelines of 20 July 2026 cite two advertising examples: a generated video showing a celebrity in a promotional context, and a generated image of a product in an ad or on packaging that could mislead consumers about its real appearance.

In other words, a fully synthetic photorealistic packshot must be labelled. A dragon flying over the volcanoes of the Auvergne need not be.

And, it bears emphasising, there is no advertising exemption. The lighter regime for evidently artistic or satirical works requires the creative nature to be obvious to the people exposed to it. As soon as content mixes the creative and the informative, the informative prevails: a video showing a synthetic influencer testing a real sponsored product does not qualify. The ARPP puts it bluntly: no exemption is provided for advertising.

The provider’s technical marking is not enough. Standards such as C2PA and invisible watermarks meet the provider’s obligation, not the advertiser’s. Information intended for the public must be perceptible without needing a tool, a specific action or metadata extraction.

Hyperrealism therefore now comes at a cost. The more a creative aims for photorealism, the more it falls within scope; a clearly stylised art direction falls outside it. This trade-off needs to be made at the brief stage. The ARPP grid combines three criteria: the level of AI intervention, the degree of realism and the impact on perceived authenticity.

Adding the “made with AI” notice: where and when?

Article 50(5) requires clear and distinguishable information at the latest at the time of first interaction or exposure: a notice buried in a footer, a menu or the terms and conditions is not enough. Three practical benchmarks:

• at the start of the video for a spot;
• at the start of the publication for a text;
• at the moment the content is encountered while scrolling a social network.

One clarification: when audiences may reasonably miss the beginning of the content, a notice placed at the start must be repeated further on.

No format is mandated: visual, audio, label, disclaimer or embedded notice, the choice depends on the audience and the medium. The guidelines set no minimum display time and no specific rule for thumbnails or looping content. In short, the criterion to remember is functional: the information must not be easy to miss under normal viewing conditions.

In an advertising context, the challenge of course lies in keeping the label intact throughout the production chain. The deployer must take steps to ensure the label remains visible at first exposure, notably through contractual terms with its distribution partners. A label embedded in a 16:9 master has to survive vertical cropping, format adaptations, programmatic trafficking and platform compression.

Note: hosting providers, platforms and broadcasters that merely transmit third-party content are not deployers. The burden does not shift to the media owner: it stays with the advertiser.

Advertiser or agency: who labels?

It is the question that comes up in every production meeting. To answer it, Article 50 distinguishes two roles:

• The provider designs the system and marks its outputs in a machine-readable format.
• The deployer visibly discloses content that constitutes a deep fake. The deployer is therefore whoever decides on the use of the AI system and controls how it is used.

The guidelines clarify how responsibility is split. A company whose employees produce content with an AI system is a deployer, without its staff becoming deployers individually. A legal entity also remains a deployer when it uses service providers acting on its behalf. Conversely, a company that hires an agency without deciding or controlling whether and how AI is used does not qualify as a deployer.

In short, the more closely an advertiser steers its agency’s use of AI, the more it takes on the deployer’s responsibility. This is an awkward position, since overseeing production remains best practice, both creatively and legally. The right approach is to allocate the execution of labelling by contract while taking on the regulatory responsibility. The scope also extends beyond the EU: the ARPP points out that an operator established outside the EU can qualify as a deployer as soon as the content is distributed there.

Let’s also bear in mind that labelling is not a shield. Complying with Article 50 makes neither the use of the system nor its output lawful. Content can be labelled and still be prohibited under Article 5, or breach data protection law, image rights, intellectual property or trademark law.

Four off-the-radar issues not to underestimate

Four types of system usually slip past marketing teams’ attention because they do not look like advertising compliance issues. News coverage of the AI Act has tended to leave them in the shadows, even though they also fall under the marketing department’s responsibility.

Chatbots and agentic commerce

A system that interacts directly with a natural person must inform them of its artificial nature, unless this is obvious from the context. The newest point concerns transactional agents: when an agent books, negotiates, concludes a contract or makes a purchase, it must disclose its artificial nature and the person on whose behalf it is acting. This is an architectural constraint for any automated prospecting setup. Formally, the obligation rests with the system’s provider.

Job ad targeting

Algorithms that place recruitment ads will be classified as high-risk when they determine who sees the ad and, in effect, exclude certain candidates. An advertiser running a programmatic employer branding campaign is affected, even though its media team has no reason to see itself as an HR player. As this topic falls under the “high-risk” Annex III, applicable from 2 December 2027, there is still time. The inventory work should start now, however, because the obligation requires obtaining documentation from the provider that does not yet exist.

Emotion measurement

Emotion recognition and biometric categorisation based on sensitive attributes are listed in Annex III: neuromarketing, creative testing through facial analysis, audience measurement in digital out-of-home. The high-risk deadline is 2 December 2027, but Article 50(3) already requires informing the people exposed. Note: in the workplace and in education, emotion recognition has been outright prohibited since February 2025, which rules out any creative testing on a panel of employees.

Political and public-interest advertising

Systems that optimise the targeting and delivery of political ads will also join Annex III on 2 December 2027. But most of the framework is already in force: Regulation (EU) 2024/900 on the transparency and targeting of political advertising has applied since 10 October 2025. Its effects are already being felt, as Meta has stopped running political and social issue ads in the EU, to the detriment of associations and local authorities in particular. The regulation’s scope goes well beyond political parties: an advocacy campaign, a stance on a social issue or a public-interest message may fall within it.

The AI Act does not stand alone

The AI Act adds another layer to the regulatory framework for advertising campaigns:

• The GDPR continues to govern profiling and requires informing people whose data feeds AI systems.
• The DSA requires platforms to provide ad-by-ad transparency and ad repositories.
• Regulation 2024/900 on political advertising has applied since October 2025.
• A Digital Fairness Act is on the Commission’s 2026 work programme, targeting manipulative interfaces, personalisation and personalised pricing.

It is probably this last text, more than the AI Act, that will restrict the core of targeting practices: where the AI Act requires transparency about creatives, the Digital Fairness Act will go after the mechanics.

Further reading: GDPR cookie audit: the DPO’s practical guide to detecting and fixing drift