An executive speaking five languages in a video, a virtual customer sharing her experience, an invented beach used to sell a holiday: generative AI has made plausible fabrication part of the communications toolkit. With the AI Act deadline of August 2, 2026 approaching, the question is changing. It is no longer just about whether a creative execution works, but whether the public understands what it is looking at. Transparency is becoming part of the message, not simply an addition in the small print.
This analysis takes a September 2026 perspective, drawing on the European regulation adopted in 2024 and information available through August 2025. It examines the expected consequences of the statutory timetable, without assuming any subsequent amendments or implementation details that may since have been clarified.
One tool, several obligations
Article 50 of the European regulation on artificial intelligence does not require an indiscriminate “created by AI” label on every piece of content. It distinguishes between several situations and allocates responsibilities between system providers and the organisations using them. For a brand, this framework matters: buying a compliant tool does not necessarily make a campaign produced with it compliant.
On the provider side, systems generating synthetic content — audio, images, video or text — must enable their outputs to be marked in a machine-readable format and detected as artificially generated or manipulated. This obligation takes technical feasibility into account. In particular, it provides an exception for functions that assist with standard editing or do not substantially alter the input data or its meaning.
For professional users, known as “deployers”, two categories are directly relevant to communications: images, audio or video constituting deepfakes, and certain texts published to inform the public on matters of public interest. Technical marking and visible disclosure therefore serve different purposes. Invisible metadata alone does not replace the explanation owed to the viewer.
Deepfakes: beyond a celebrity’s face
In the public imagination, a deepfake is still a doctored political video or a fraudulent impersonation. The European definition goes further: it covers AI-generated or manipulated content that resembles existing people, objects, places, entities or events and would falsely appear authentic or truthful.
An advertisement showing a real hotel with an invented swimming pool may therefore raise a different issue from an obviously fantastical animation. The same vigilance is needed for a cloned spokesperson’s voice or a sequence making an executive say words they never recorded. Classification depends on the content and how it is presented, not simply on the software used.
Conversely, not every synthetic advertising creation is automatically a deepfake. An abstract world or an openly unreal mascot does not create the same impression of authenticity. That does not remove the obligation to comply with other rules, particularly those on misleading commercial practices. A fictional customer presented as a genuine testimonial poses problems far beyond its AI label.
Creativity is not a free pass
Article 50 provides tailored rules when content forms part of an evidently artistic, creative, satirical, fictional or similar work or programme. The disclosure obligation remains, but may be fulfilled in an appropriate manner that does not hamper the display or enjoyment of the work. This is not a blanket exemption for advertising. Invoking a campaign’s creativity is not enough to remove the duty to inform.
Public interest also extends to brand communications
Another sensitive area is text generated or manipulated by AI and published for the purpose of informing the public on matters of public interest. The obligation is not confined to the media. A company explaining industrial pollution, a health issue or the social consequences of a site closure may find itself in this territory.
The regulation does, however, provide an exception where the content has undergone human review or editorial control and a natural or legal person assumes editorial responsibility for its publication. For communications teams, the challenge is therefore as much organisational as visual: who checks the facts, who approves the text, and who is accountable for it?
Automatic approval or a rushed read-through should not become a pretext for circumventing the spirit of the rules. Without inventing an additional mandatory procedure, an organisation has an interest in retaining evidence of a genuine review. In sensitive communications, being able to explain how the text was checked also matters for credibility.
Transparency starts with the brief
The most tangible change should take place before production. Today, an agency may deliver a film in which only some shots are synthetic, a studio may add a cloned voice, and a platform may then produce several versions. If no one documents these operations, the person responsible for publication inherits an asset without knowing how it was made.
A practical approach is to incorporate four questions into every brief:
- Which elements will be generated or modified by AI, and by which service provider?
- Could the public mistake these elements for an authentic representation of an existing person, place or event?
- Does the text aim to inform people about a matter of public interest, and what editorial oversight is planned?
- How will the transparency disclosure remain accessible in every version?
This last point is crucial. A legible disclosure in a long-form film may disappear when it is cropped into a vertical video. A clarification placed beneath a post may be lost when it is reshared. Article 50 requires clear, identifiable information, provided no later than the first interaction or exposure, in compliance with applicable accessibility requirements. Transparency must therefore be designed for the format people actually see.
Informing without normalising doubt
Then comes the choice of words. “AI-generated visual” does not explain the same thing as “synthetic voice reproducing that of our spokesperson”. A precise disclosure can better describe the transformation, without guaranteeing compliance on its own. The aim is not to recount the entire production chain, but to prevent a false interpretation of the content.
Provenance mechanisms, particularly those developed around the C2PA standard, can complement this approach. They document an origin or modifications; they do not prove that an advertising claim is true. Nor does the absence of a disclosure constitute a certificate of authenticity. For brands, the risk would be to treat the label as a substitute for verification.
What next? The next step may be less spectacular than AI demonstrations: better-drafted contracts, documented approvals and disclosures tested with the public. The best-prepared brands will not necessarily be those that display “AI” most often, but those that can explain exactly what has been fabricated — and take responsibility for what they publish.