Article 50(4) of the EU AI Act (“Act”) requires “deployers” of AI systems to clearly disclose that content has been artificially generated or manipulated in two distinct scenarios:
- where image, audio or video content qualifies as a “deepfake” (a broadly defined concept, explored further below) which is likely to capture a raft of AI-generated advertising and creative visual assets; or
- where AI generated or manipulated text is published with the purpose of informing the public on matters of public interest and has not undergone human review or editorial control. Routine AI-drafted advertising copy will generally sit outside this second scenario, provided it does not include claims relating to matters such as health, consumer safety or sustainability.
The reason for these new labelling rules is so that audiences should be able to tell fact from fabrication, rather than being quietly deceived that a photorealistic AI creation is real or authentic. However, the practical assessment as to who is the “deployer” of AI in a typical advertising chain, and what may be a “deep fake” under the Act, is an objective test to be determined on the facts.
Who’s caught by this obligation?
More people than you might expect, and not just those based in the EU. Even though the Act is a piece of European legislation (which has not been implemented in the UK), it has “extra-territorial” reach, meaning the AI labelling rules apply to both:
- deployers established or based within the EU; and
- deployers based in third countries outside the EU (e.g. in the UK or the US) where that deployer itself foresees, directs or authorises distribution of AI content into the EU, including by posting "deepfakes on the globally accessible internet".
It is worth noting that non-EU deployers are not expected to be caught where their content only reaches EU audiences through channels that are “unforeseeable and outside their control”. So, whilst agencies and advertisers running campaigns with obvious EU reach should assume the rules apply (such as via the publicly accessible internet, even where a campaign is not formally targeted at the EU), a campaign geo-targeted to the UK only that reaches the EU only incidentally and unforeseeably (e.g. shared privately via WhatsApp) is unlikely to be caught.
Who are “deployers” and why is this relevant to advertisers and agencies?
“Deployers” are those who use AI systems under their “authority” and who control how the AI system is used. “Authority” over an AI system is understood broadly. It means assuming responsibility for the decision to deploy the system and for the manner of its actual use (including its outputs), but it does not require hands-on technical control over the tool itself. It is enough that a party decides for what purposes and how the AI system is used, even across decentralised workflows or multi-party supply chains.
Applying that test to an advertising campaign, it is perfectly conceivable for both the advertiser and the agency to be “deployers” within the same chain, each in a different respect. For example, a client may set the parameters and rules governing AI use across a campaign (e.g. approving which use cases are permitted, imposing brand safety and content restrictions, and signing off final creative), while the agency retains day-to-day control and discretion over which specific AI tools are used and how they are applied during the creative process. Both these roles can amount to “authority” over the AI system’s use, even though neither party has complete, exclusive control. In this scenario, compliance is very likely to be a joint responsibility, not something you can quietly leave to the other party.
What is a “deepfake”?
AI content needs to satisfy a two-stage test to be considered a “deepfake”. The content must:
- resemble existing people, objects, places, entities or events; and
- it must falsely appear to a person to be authentic or truthful.
Both boxes need to be ticked. The guidance is clear that genuinely minor or technical AI interventions, such as background tidying; lighting and colour adjustments; cosmetic touch-ups; or aesthetic background replacements and product re-scaling, are unlikely to meet the test, because they have only a minor impact on how authentic or truthful the content appears to its audience.
Interestingly, official guidance on the Act suggests “existing” does not mean the subject must be a real, identifiable person, object or event: it is enough that the simulated subject resembles someone or something “that exists, could plausibly exist, or could plausibly have existed in reality”. Only genuinely fantastical content that defies the laws of nature or physics (think dragons, or elephants driving cars) falls outside this, because it has no real potential to mislead. Because this first limb is presented so broadly, in practice this means the assessment of whether something constitutes a “deep fake” requiring a label will often turn on the second limb of the test: does the content falsely appear to a person to be authentic or truthful?
The second limb is assessed objectively and it does not require any intention on the part of the advertiser or agency to deceive. Context will be crucial when it comes to assessing whether the “falsely appear” test is met: relevant factors include the level of resemblance to the real subject; the substantive message the content conveys; the intended and reasonably foreseeable context in which the content will be deployed; the environment in which it is presented; and the composition of the intended and reasonably foreseeable audience and their expectations, including whether children, older people or other groups who may be more easily misled are likely to be exposed.
This is particularly true in an advertising context, where the guidance is clear that the focus belongs on the advertised product, not the scene dressed around it: “AI-powered colour correction, background extensions of existing content, adjustments or replacements of backgrounds for clearly aesthetic purposes, compositions and arrangements of existing products, or re-scaling of images applied in product advertisements or packaging is likely to have only a minor impact on a person’s perception of the authenticity and truthfulness of the advertisement and the product.” Put simply, an AI-generated backdrop that merely sets a scene – whether stylised or aiming for realism – is unlikely to require labelling, provided it does not affect how the audience perceives the authenticity of the actual product, person or event the advert focuses on or attempts to sell. The position may be different where the AI use touches the very item being sold or promoted, for example a product rendered as better, bigger or different than reality.
On these grounds, a fox driving a motorboat won’t meet the test, because one doesn’t normally see foxes at the helm of a yacht in the waters off Cannes.
But a hyper-realistic AI-generated version of a celebrity, or a product rendered suspiciously better than its real-life counterpart, almost certainly will. The new regime creates a real danger. Fraudulent deep-fakes promoting financial scams, like this one of financial journalist Martin Lewis, will not carry any disclosures because they are fakes propagated by fraudsters. But as disclosures become more widespread, there is a risk that gullible consumers are more likely to be tricked into believing a fake is a genuine by virtue of the absence of any disclosure label.
As above, context will be key: an AI-generated cartoon of a historical event is unlikely to fool anyone, whereas a synthetic influencer (resembling a real person) convincingly promoting a product might. There is extensive (and not always clear) guidance from the EU Commission around both parts of this test and ultimately, whether something is a “deep fake” or not, will depend on the facts and will need to be assessed on a case-by-case basis.
If your content meets the “deepfake” test, how should you label it?
The disclosure must be clear, distinguishable and understandable to the audience, factoring in whether children, older people or those with accessibility needs might be watching. It needs to appear when the content first reaches its audience, not buried in the small print or hidden in metadata nobody will ever see. There’s no single mandated format, but the European Commission has recommended an EU icon (a simple “AI” mark, with “generated” or “modified” as appropriate) in its Code of Practice on Article 50, which deployers may use to help demonstrate compliance (though adherence to the Code is not, of itself, conclusive evidence of compliance, and use of the EU icon is not mandatory). It is worth noting that AI deepfakes created before 2 August 2026 do not need to be marked or labelled retroactively but it would be considered good practice to do so.
Advertisers may have a preference on what AI label they want to include (and where) in campaign assets, but advertisers and agencies will need to make sure that the AI label is noticeable and clear while also preserving the quality of the creative work. Most online platforms (like Instagram and TikTok) will have built-in tools to label AI generated or manipulated content (either voluntarily by the user disclosing that AI has been used or automatically through the platform’s built-in AI detection systems). However, if the platform label is invisible (i.e. a machine-readable marking that AI tool “providers” embed in content to satisfy their own disclosure duty or if it is not prominently visible on or beneath the post), deployers will not be able to rely on this to discharge their own disclosure obligations since that marking is not immediately clear and distinguishable to the natural persons exposed to the content; a human-perceivable and prominent label is still required. Even if the platform-applied AI label is prominent on the post and visible to a human user, it is still the advertiser and the agency’s responsibility to consider if AI disclosure is required for that content and if so, if the AI label is compliant with the Act’s requirements.
Real word consequences: not just a slap on the wrist.
Non-compliant deployers face fines of up to €15 million or 3% of global annual turnover, whichever is higher. It is not clear at this stage how this will be enforced in the context of ad campaigns, but the industry has already shown concern regarding the breadth of this disclosure obligation and the possibility of it applying to routine commercial imagery in ad campaigns, diluting the value of those campaigns to consumers.
It’s also worth remembering that labelling a deep fake does not, on its own, make the content lawful or exempt it from other rules. The AI Act’s transparency obligations sit alongside, and do not override, other applicable regimes such as data protection law (where a real person is depicted), intellectual property and personality rights, and, in the UK, existing advertising self-regulation and consumer protection law. Compliant labelling addresses the AI Act’s disclosure requirement, not the underlying lawfulness of the advertisement.
What does this mean for advertisers and agencies?
Agencies are increasingly leaning on AI in campaigns for backgrounds, synthetic talent, and product renders to help with cost pressures; and advertisers are encouraging agencies to do so to create better campaigns for less.
Both agency and advertiser need to work out, asset by asset, whether they are “deployers” within the meaning of the Act and whether the two-stage test for a “deep fake” is met by the content in question. This will often come down to context – in particular, whether the AI-generated or manipulated element affects the audience’s perception of the authenticity of the actual product, person or event being advertised, rather than merely dressing the scene around it.
Where in scope, they should agree early in the creative process what label to apply, and where it features in the asset, making sure that the label is present in the asset at the time of publication or broadcast, rather than after the campaign has gone live across Europe.
Given there is a likelihood that both advertiser and agency could be treated as independent “deployers”, this is not a problem that can be quietly outsourced down the supply chain and the advertiser and agency will need to agree on how to allocate their AI labelling responsibilities and associated risk in contracts (until industry standard positions are established).
The safest approach is to build the AI disclosure conversation early into the creative process, rather than treating it as afterthought at the eleventh hour when both advertiser and agency are approving final versions of the campaign ready for publication or broadcast.
See here for more information.
