A month after the EU's content-transparency rules became enforceable, the Dutch advertising industry has produced what may be the first sector-specific reading aimed at creative agencies. VIA Nederland, formally the Vereniging In Advertising, published guidance on 5 September setting out four practical tests agencies can apply to decide whether a campaign needs an AI label, and how that label should look.

The advice matters because Regulation (EU) 2024/1689, the AI Act, has been binding since 1 August 2024, with Article 50's transparency obligations enforceable from 2 August 2026. The European Commission issued its final guidelines on 20 July, thirteen days before the compliance date. Non-compliance can carry fines of up to 3% of worldwide annual turnover.

The question is not which tool an agency used

VIA Nederland's central argument is that disclosure hinges on the agency's role and the use case, not on the name of the software involved. A script drafted with ChatGPT and a synthetic voice created in ElevenLabs trigger different obligations; a Midjourney illustration used as a mood board and the same illustration used as the lead image of a campaign trigger different obligations again.

Running underneath that logic is the AI Act's split between providers, who place an AI system on the market under their own brand, and deployers, who use such a system under their own authority. VIA Nederland treats most agencies as deployers, since they typically work with existing models such as ChatGPT, Midjourney or Adobe Firefly. The classification changes once an agency builds its own application on top of an underlying model: at that point it may have to assess whether it has also become a provider for that specific tool.

That distinction shapes the rest of the framework. Providers carry machine-readable marking duties, watermarks and metadata that make synthetic content technically detectable. Deployers carry human-detectable labelling duties, the visible or audible notice a viewer or listener must actually perceive.

When a chatbot needs an introduction

The first scenario addresses AI systems that communicate directly with people: a brand chatbot, a product adviser, a voice assistant. The provider must ensure users know they are talking to a machine, unless that fact is already obvious from the context. VIA Nederland gives the example of a technical assistant embedded in an engineering tool, which would naturally read as artificial, against a consumer-facing chatbot styled to look like a human employee, which would not.

The notice must arrive at the moment of first contact, in a short text message for a chatbot or a spoken line for a voice assistant. Burying the disclosure in terms and conditions does not satisfy the requirement, the guidance says. For agencies, the duty bites mainly when they build interactive tools for clients rather than reselling an off-the-shelf product.

Images, voices and the deepfake test

Synthetic voices, virtual presenters, AI-generated environments and digitally altered footage are the area where agencies face the most production decisions. VIA Nederland applies the AI Act's deepfake definition as the operative test: AI-generated or manipulated content resembling existing people, objects, places or events in a way that could create a false impression of authenticity.

A realistic AI-generated person who could be mistaken for a real actor falls inside the definition. So does the cloning of an existing voice. A minor technical adjustment to colour grading does not, because the audience would not be misled about authenticity. Fictional scenes and clearly stylised effects fall outside too, because viewers do not expect them to depict reality.

The guidance spends unusual care on the artistic-content carve-out. Under the AI Act, creative, artistic, satirical and fictional works can present a deepfake label less prominently. VIA Nederland warns that this lighter regime cannot be invoked once the commercial message dominates, which in practice means almost every advertising campaign. Bird & Bird's analysis from 11 June 2026 reached the same reading: a manipulated video aimed at selling a product is not a creative work for these purposes.

Italy's data protection authority has tested the limits in a different direction. In its decision on synthetic depictions of the journalist Enrico Mentana, the regulator ordered broadcaster Mediaset to remove the material within 30 days, finding that a label did not cure the underlying harm of placing synthetic performance inside a real professional context.

Where a label is required, it has to be perceptible. Metadata on its own is not enough; the viewer or listener must actually see or hear the disclosure. The Commission's icon set, three designs with four visual variations available in SVG and PNG, is optional. Agencies may design their own, as long as the underlying requirement of clear disclosure at first exposure is met.

Text, but only sometimes

For AI-generated or AI-manipulated text, the deployer obligation narrows sharply. It only applies when the text is published to inform the public on matters of general interest, covering politics, public health, public safety, fundamental rights or other topics relevant to public debate. A human review exception applies where a person with relevant knowledge has substantively assessed, adjusted or could reject the content. VIA Nederland recommends documenting that intervention as part of the workflow; a spell check or a handful of rewritten sentences does not count.

Building your own tool changes the calculation

The fourth scenario is the one VIA Nederland flags as least common but most consequential. Agencies that develop their own AI systems, or applications layered on existing models, may qualify as providers for those tools and inherit provider-level obligations: marking output in a machine-readable way as AI-generated, separately from any visible label a deployer downstream might add.

In practice most agencies inherit that marking for free. The Code of Practice on Transparency of AI-Generated Content, which the Commission finalised earlier this year, requires signatories to apply at least two machine-readable layers, digitally signed and time-stamped metadata alongside an imperceptible watermark. Google, Meta and Anthropic signed in July 2026, covering most of the models Dutch agencies use day to day. An agency building a custom generative layer still has to consider whether its own output needs separate marking.

Five questions, run early

VIA Nederland's recommendation is to fold the transparency check into the creative process rather than treat it as a legal review the night before launch. The guidance sets out five questions for any campaign involving AI: identify the agency's role; classify the application, whether it talks to consumers, generates or manipulates imagery, or publishes text on matters of general interest; check whether a viewer could be misled about authenticity; ensure human review is documented; and confirm how the disclosure will appear without disrupting the creative work.

Organisations

VIA Nederland · European Commission · Bird & Bird · Google · Meta · Anthropic