AI Act in advertising: when should AI-generated content be labelled?

Two obligations that are easily confused

The provider of a generative AI tool has duties concerning machine-readable marking and detection of its outputs. The professional user of an AI system, known as a deployer, has separate duties towards people exposed to certain content. An agency using a generator should usually start with this second perspective. Commissioning an image does not, by itself, make the client a provider of an AI system.

Technical information stored within a file does not replace an intelligible disclosure to the audience where one is required. Nor does the transitional period ending on 2 December 2026 generally postpone the labelling of advertisements. That period concerns providers’ duties under Article 50(2) for systems placed on the market before 2 August 2026. It does not give agencies extra time to start disclosing deepfakes.

Retouching, product visualisations and content that appears real

Under the AI Act, a deepfake is AI-generated or manipulated image, audio or video content resembling existing people, objects, places, entities or events that would falsely appear authentic or truthful. The issue extends beyond putting a celebrity’s face into someone else’s video. A campaign may also suggest that an event actually happened or that a product really looks or performs as depicted.

Removing noise or adjusting white balance while preserving the meaning of a photograph is different from adding a function that the product does not have. The boundary does not depend on the software’s name or a button labelled “retouch”. Recolouring clothing may accurately show an available variation, or it may depict a version the client does not sell. In the latter case, the claim conveyed by the image needs correcting. An AI label alone does not resolve it.

With a realistic mockup, the agency should establish whether the audience will see a proposed arrangement or what appears to be a photograph of completed work. A generated interior visualisation and a photograph presented as a finished apartment make different promises. Similarly, a product in a fantastical setting differs from footage presented as a record of a real test. The image, caption, publication context and audience expectations all matter. Photorealism warrants closer assessment but does not determine the answer on its own.

The standard-editing exception in Article 50(2) concerns the provider’s technical marking obligations. It should not be treated as a general exemption from deepfake disclosure for everything described as post-production in a project brief.

Synthetic narration, voice cloning and virtual influencers

Using a synthetic narrator without presenting it as a particular person or as a genuine recording of that person’s statement does not, on its own, settle whether the result is a deepfake. Content and presentation still need assessing. Neutral narration of instructions differs from a voice suggesting that a real expert personally endorses a product. With a recognisable voice clone, the agency must also consider whether the audience may believe that the individual actually recorded those words.

The fact that an advertising character does not exist is not an automatic defence. The European Commission’s interpretation also encompasses realistic characters who could plausibly exist. An avatar presented as a real customer, with a name, purchase history and account of the results, therefore needs particular attention. Calling it a “virtual influencer” in the agency’s production notes does not inform the audience.

Disclosure of AI use and the truthfulness of an endorsement are separate issues. If the character describes a purchasing experience that nobody had, an “AI-generated” label does not make the testimonial lawful. The concept may need to change: the character could be presented as a fictional narrator, with claims about personal product use removed. Where a real creator is involved, the agency must separately assess permission to use their image, voice and source material.

Advertising copy and text about health or the environment

Article 50 does not generally require a label on every AI-assisted slogan or product description. A separate rule covers text published for the purpose of informing the public on matters of public interest. A short promotional message and a PR article about an ingredient’s health effects may therefore call for different decisions. Consumer safety and environmental protection can also fall within this category. Describing a piece as “marketing content” does not determine the purpose of its publication.

The exception for such text requires genuine human review or editorial control and a natural or legal person holding editorial responsibility for publication. In practice, someone should examine the substance of the claims, the data and the reliability of the sources, with the ability to change or reject the text. Checking punctuation or approving the page design is insufficient. Keeping a record of who reviewed the relevant version and the basis for approving significant claims is useful.

This exception concerns text. It does not remove the duty to disclose a deepfake in an accompanying video simply because the client approved the script. A publication combining an article, a synthetic statement and an illustration may require a separate assessment of each element.

How should the disclosure reach the audience?

Required information must be clear, distinguishable and available no later than the audience’s first exposure to the content. Applicable accessibility requirements also matter. A label buried among hashtags, available only after expanding a caption or stored solely in the file’s metadata does not reliably inform the audience in time. For a standalone audio advertisement, the disclosure needs to be audible. Text below a player will not reach someone hearing the recording elsewhere.

The wording should reflect what was actually generated or altered. “Image and voice generated using AI” is more informative than a vague “AI powered”. Where an authentic recording has been manipulated, the disclosure should communicate that manipulation instead of suggesting only technical editing assistance. No single phrase makes every asset compliant: placement, contrast, display duration and format also matter.

Advertising disclosure and AI disclosure answer different questions. The first identifies the commercial nature of the message; the second explains the artificial origin or alteration of particular content. Where both are needed, an advertising label does not replace AI disclosure, and AI disclosure does not replace an advertising label. This distinction also matters for influencer campaigns directed at Polish consumers.

Does creative advertising qualify for an exception?

The AI Act recognises deepfakes forming part of evidently artistic, creative, satirical, fictional or similar works or programmes. It allows disclosure of their presence in an appropriate manner that does not hamper the display or enjoyment of the work. This adjusts how the duty is fulfilled; it does not remove the duty altogether.

An agency therefore cannot rely simply on the argument that all advertising is creative work. The particular asset and the way its audience will understand it need assessing. An openly fantastical story differs from a video resembling a spontaneous customer review. The provision should not be treated as automatic permission to postpone disclosure until the closing credits either: the first-exposure requirement remains relevant.

Agency, freelancer and client: who is responsible?

AI Act roles depend on the actual use of the system and the authority under which it is used. An employee acting under an agency’s direction does not thereby become a separate deployer. For a freelancer, the relevant question is whether they use the tool independently or on the agency’s behalf and under its control. A client receiving a finished file does not automatically become a deployer merely because it publishes the output.

The agreement should nevertheless identify who prepares the disclosure, checks it before release and preserves it in later adaptations. Assigning a task to the client does not, by itself, change the agency’s statutory role. The client also needs enough information to avoid removing a disclosure while shortening a video or producing a new format. This allocation should align with the approval process in the marketing agency agreement.

A draft presentation or moodboard is not automatically outside the assessment. What is shown, to whom and whether synthetic material may be understood as authentic all matter. Narrow guidance on business-to-business uses in the context of providers’ technical output marking does not create a general exemption from deepfake disclosure. Even at presentation stage, it is sensible to distinguish an AI concept from material documenting completed work.

What should accompany the client deliverables?

At handover, the client needs a short explanation of which elements were generated or altered using AI, which version was approved and where and how disclosure must appear. That explanation should identify particular files and channels. The client should know what it may change in the material and which required AI disclosure must remain when the video is shortened or adapted to another format.

The same arrangements should identify responsibility for publication and subsequent adaptations, licence restrictions, and the scope of permissions concerning image and voice. There should also be a response process after launch: who reports missing disclosure, who can pause distribution and who provides a corrected file. Broader rules on rights, data and confidentiality are discussed in Generative AI in a Polish marketing agency.

The team needs to be able to apply these arrangements. Article 4 of the AI Act currently requires measures supporting the development of AI literacy among people using AI on an organisation’s behalf; it does not require a guaranteed level of knowledge for each individual. Agency training can use the types of material the team actually produces: the difference between correction and a changed message, when a character appears to be a genuine customer, and why approving text does not remove the need to disclose a deepfake in a video.

An AI label does not replace advertising review

A label does not grant rights to someone else’s face or voice, expand a licence or justify untrue product claims. The advertising message and the rights in the material need separate review. Where a creator is involved, permissions and permitted modifications should be addressed in the influencer agreement.

The appropriate response may be to add disclosure, but it may also involve rewriting the script, changing an image or removing a false endorsement. That decision requires assessment of the actual asset.