Not every commercial made with Midjourney needs a label, and not every script drafted with ChatGPT does either. According to VIA Nederland, the Dutch industry association for advertising, what actually triggers a disclosure duty under the European Union's AI Act depends on the application, the agency's legal role and the type of content being produced, a distinction the group laid out today in guidance aimed squarely at creative agencies.

In Short

The European Union now requires companies to say when they used artificial intelligence to make certain kinds of ads, chatbots or public-interest content, but the rule does not cover everything made with AI tools. This matters for advertising agencies, because getting the scope wrong either means slapping labels on harmless work or missing a real legal requirement that can carry fines. VIA Nederland's guidance gives agencies four situations to check for every campaign, so the labeling decision becomes a routine step rather than a last-minute scramble before launch.

A framework built around role, not tool

VIA Nederland, formally Verenigd In Advertising, published its guidance as a practical walkthrough of Article 50 of the AI Act, the transparency chapter that became legally applicable across the European Union on August 2, 2026. The EU AI content rules force publishers to label or risk 3% of turnover coverage from PPC Land detailed how that date followed a two-year runway after Regulation (EU) 2024/1689 entered into force on August 1, 2024, and how the Commission converted months of draft text into settled guidance on July 20, 2026, less than two weeks before the compliance date arrived.

Rather than restating the regulation's structure, VIA Nederland's guidance walks through four activities a creative agency is likely to perform on any given day: building a chatbot, generating an image or video, drafting text, or building an AI tool on top of an existing model. For each, it sets out whether a disclosure duty attaches and, if so, what form the disclosure has to take.

The starting distinction sits underneath everything else. Under the AI Act, a provider is an organization that develops an AI system, or has one developed, and places it on the market under its own name or brand. A deployer, called a gebruiksverantwoordelijke in the Dutch text, uses an AI system under its own authority for professional purposes. According to VIA Nederland, most creative agencies fall into the deployer category, because they typically use existing systems such as ChatGPT, Midjourney or Adobe Firefly rather than building their own models. The line blurs when an agency builds a custom application on top of an existing model, at which point each application may need separate assessment for whether the agency has also become a provider.

That split runs through the entire Article 50 framework. Providers of generative systems face marking duties aimed at making synthetic output machine-detectable; deployers face labeling duties aimed at making it human-detectable. VIA Nederland's guidance treats that distinction as the organizing principle for everything that follows.

Chatbots and brand assistants: say it is AI

The first scenario VIA Nederland addresses is the one most agencies will recognize immediately: a chatbot answering questions on behalf of a brand, or an AI brand assistant offering product advice. According to VIA Nederland, when an AI system communicates directly with people, the provider of that system must ensure users know they are talking to AI. An exception applies when this is already obvious to the user, and the guidance stresses that context and circumstances determine what counts as obvious. A technical AI assistant embedded in an engineering tool will more readily read as artificial than a consumer-facing chatbot styled to resemble a human employee.

On how that disclosure has to look, VIA Nederland is specific: the notice has to be clear at the moment of first contact. A chatbot can use a short text notice, a voice assistant a spoken one. A mention buried only in terms and conditions, or one a user would have to search for, does not satisfy the requirement, according to the guidance.

For agencies, this duty becomes relevant mainly when they build interactive AI tools for clients rather than deploying an off-the-shelf assistant. Where an existing solution is used instead, the guidance advises agencies to determine which party in the chain carries the disclosure responsibility, since the provider and the deployer of an interactive system are not always the same organization.

Images, audio and video: the deepfake test

The second and most detailed section of the guidance covers image, audio and video content, the terrain where creative agencies spend the most production time and carry the most exposure. Synthetic voices, virtual people, generated environments and AI-modified footage all fall under this heading, and the question VIA Nederland poses is when any of it needs an AI label.

The operative test, according to the guidance, is whether the finished content qualifies as a deepfake under the AI Act's definition. That happens when AI-generated or manipulated image, audio or video content resembles existing people, objects, places, entities or events in a way that could wrongly create the impression of being authentic. A further detail matters: an AI-generated person does not have to resemble a specific known individual to trigger the definition, since what counts is whether the created person could be mistaken for a real one. In substance, the guidance frames the question as whether a viewer could be deceived about the authenticity of the content.

The guidance gives concrete production scenarios. A commercial featuring realistic AI-generated people who could be mistaken for real actors falls inside the definition, and so can face replacement or the cloning of an existing voice. But not every AI-assisted edit automatically becomes a deepfake, according to VIA Nederland, since audience context and expectations factor into the assessment. Special effects or clearly fictional scenes can make it obvious to viewers that what they see does not depict reality, and a minor technical adjustment to color grading or lighting does not need disclosure either. The mere fact that AI featured somewhere in the production process does not, by itself, turn the finished output into a deepfake.

Creative, artistic, satirical and fictional works receive a specific carve-out under the AI Act, and VIA Nederland's guidance walks through it carefully because it is the provision most likely to be misapplied by agencies hoping it covers advertising. Where AI is used in such a work and the result qualifies as a deepfake, a transparency duty still applies, but the label may be presented less prominently, so the disclosure need not appear as conspicuously in view or audio as it would elsewhere. Where the commercial message dominates, however, that lighter regime cannot be invoked, and the guidance states plainly that this exception will therefore not apply to creative advertising messages.

That reading lines up with how the Commission's own guidelines have been interpreted elsewhere. According to a Bird & Bird analysis published June 11, 2026 and cited in PPC Land's coverage of the EU publishes free AI labelling icons ahead of August 2026 deadline, an AI-manipulated video simulating people in an AI-generated scene depicting product use, aimed at persuading viewers to buy that product, does not qualify for the lighter artistic-content regime. The same draft guidelines exclude AI-generated imagery of celebrities implying involvement in activities they never took part in, unless that imagery serves a genuine fictional or satirical purpose. Italy's data protection authority reached a related conclusion in its decision on synthetic depictions of journalist Enrico Mentana, covered by PPC Land in Mediaset faces 30-day deadline after Italy bans Mentana deepfakes, where the regulator found that disclosure did not cure the underlying harm once a synthetic performance was placed inside a real person's professional setting.

On the mechanics of the label itself, VIA Nederland's guidance again sets a clear bar. Where a deepfake is present, the deployer carries a transparency duty: at the moment someone is first exposed to the content, it must be clear that it was generated or manipulated with AI. Metadata alone, or an invisible technical watermark alone, is not sufficient, according to the guidance. The person seeing or hearing the content must actually be able to perceive the disclosure, whether through a textual, visual or auditory notice, though the Commission's example icons are optional and agencies remain free to design their own as long as it meets the underlying requirements.

That icon set, three designs each with four visual variations in SVG and PNG format, is one of the more concrete deliverables to emerge from the Article 50 process. PPC Land's earlier reporting noted the icons apply to deployers of generative AI systems, including advertisers, marketing agencies, publishers and platforms operating in or serving the European Union, with placement rules requiring the label to be perceivable at the latest at first exposure, without intervening overlays.

Text: the general-interest threshold

The third scenario covers AI-generated or AI-manipulated text: social posts, blogs, advertising copy. VIA Nederland's guidance narrows the disclosure duty considerably here. A deployer only faces a transparency obligation when AI generates or manipulates text published to inform the public about a matter of general interest, such as politics, public health, public safety, fundamental rights or other developments relevant to public debate. For creative agencies, this is most likely to arise around societal or public-awareness campaigns rather than routine commercial copy.

An important exception applies when the text has undergone human review and a person or organization bears editorial responsibility for the publication. That human oversight has to be substantively real, according to the guidance: a spelling check or rewriting a handful of sentences is not enough. Someone with relevant knowledge has to be able to assess, adjust or reject the content, and the guidance recommends documenting that intervention as part of the workflow.

Building a generative tool in-house: providers face a different bar

The fourth scenario is the one VIA Nederland flags as least common but most consequential when it applies. Most creative agencies fall under the deployer obligations described above, since they use existing models rather than developing AI systems from scratch. Agencies that do develop their own systems, or applications layered on existing models, may qualify as providers for those specific tools, and providers face an additional set of duties.

Providers of AI systems generating synthetic text, audio, images or video must, in principle, mark that output in a way that is machine-readable and detectable as AI-generated, according to the guidance. This is distinct from the visible label required for deepfakes: the machine-readable mark makes the artificial origin technically detectable, while a visible label informs the public directly. Where AI performs only a supporting function in standard editing, or where the input's meaning is not materially altered, that machine-readable marking is not required.

This is the same marking regime PPC Land has tracked on the provider side of the market. The finalized Code of Practice on Transparency of AI-Generated Content requires providers to apply at least two machine-readable marking layers for most content types: digitally signed and time-stamped metadata alongside an imperceptible watermark, with free-form text over 200 tokens also requiring watermarking. That obligation sits with model and tool providers such as Google and Meta, both of which signed the voluntary Code in July 2026 according to PPC Land's reporting in Google signs EU AI code as advertisers face 3% turnover fines August 2 and Meta faces 3% turnover fines in 5 days as it signs EU AI content code. Anthropic, publisher of the Claude models, also became a signatory, as PPC Land detailed in Claude text gains invisible watermarks across 5 Anthropic products, rolling out imperceptible text watermarks and file-level provenance metadata across five products.

An agency that only uses these providers' tools inherits the benefit of that machine-readable marking without having to build it. An agency that builds its own generative layer on top of one of these models takes on a separate, provider-level obligation to consider whether its own output needs the same kind of marking, independent of whatever the underlying model already does.

A five-point routine, not a pre-launch check

VIA Nederland's practical recommendation is to fold transparency into the creative process itself rather than treating it as a legal check performed just before a campaign goes live. The guidance sets out five questions for any campaign involving AI: determine the agency's role, whether it uses an existing AI tool or offers an AI application of its own; look at the specific application, whether AI communicates with consumers, generates or manipulates image, audio or video, or produces text about matters of general interest; check image, audio and video output for whether a viewer could be misled about authenticity, meaning a possible deepfake; ensure human review is documented, recording who assesses AI output substantively and who carries responsibility for publication; and run a final check on whether AI use must be disclosed and, if so, how to do that clearly without disrupting the creative work.

That last point, doing the disclosure without disrupting the creative execution, is where the guidance's earlier discussion of the artistic-content exception becomes practically relevant again. Advertising cannot claim the lighter disclosure regime reserved for artistic and fictional works, so agencies producing deepfake-qualifying content for commercial campaigns need to plan the visible label into the creative from the start rather than treating it as an afterthought that a lighter regime might excuse.

Why this lands now

The publication date matters because of what else has happened around Article 50 compliance since the deadline passed. The Council of Europe's parallel work on AI and privacy, covered by PPC Land in Council of Europe drafts privacy rules for AI chatbots and agents, is scheduled for Bureau-level discussion in Paris on September 16 and 17, 2026, adding a second regime agencies running chatbots will need to track alongside Article 50. The EU AI Office also opened three complaint routes covering suspected breaches of the transparency duties, as reported in EU AI Office opens three complaint routes covering Google and Meta systems, so the compliance question is no longer purely theoretical.

There is also a commercial argument for getting the disclosure question right, independent of legal exposure. Research from the Interactive Advertising Bureau, published January 16, 2026 and covered by PPC Land in IAB introduces disclosure framework as Gen Z trust in AI ads plummets 19 points, found that 82 percent of advertising executives believed Gen Z and Millennial consumers felt positively about AI-generated ads, against only 45 percent of those consumers who actually did, a gap that widened from 32 percentage points in 2024 to 37 points in 2026. A separate Raptive study cited in the same reporting found that content suspected of being AI-generated cuts reader trust by close to half, with a 14 percent decline in purchase consideration once readers believed the content was synthetic, an effect echoed in AI ad labels cut click-through 31.5%, IAB framework cites NYU study.

Those figures cut against any simple assumption that compliance and commercial interest point in the same direction. Labeling something as AI-generated appears to measurably reduce trust and click-through in some studies, giving agencies a business incentive to apply labels only where legally required rather than everywhere as a blanket precaution. VIA Nederland's scenario-based approach speaks directly to that tension: over-labeling carries a cost, and under-labeling carries legal risk, so getting the line right in each case matters more than defaulting to either extreme.

The regulatory landscape around synthetic advertising content has also been shifting in ways that reinforce the deepfake-focused portion of VIA Nederland's guidance. Brazil's national data protection authority published a 163-page technical report on July 29, 2026, mapping how deepfakes are used in advertising fraud, as PPC Land reported in Brazil bans paid ads for deepfakes in final 72 hours before elections. A German court separately ruled that AI voice cloning without consent violates personality rights, ordering damages in a case covered by PPC Land in German court rules AI voice cloning violates personality rights. Neither case is Article 50 enforcement specifically, but both show regulators and courts across jurisdictions converging on concerns about synthetic media in commercial contexts, concerns VIA Nederland's deepfake test is designed to help agencies anticipate rather than discover after the fact.

What remains unresolved

VIA Nederland's guidance is careful to note where responsibility sits ambiguously. The provider-deployer line blurs whenever an agency builds a custom application on top of an existing model, and the guidance offers no bright-line test for exactly when that threshold is crossed, only the instruction to assess each application individually. On interactive systems such as chatbots, the guidance similarly acknowledges that when an existing solution is deployed rather than built in-house, determining which party in the supply chain carries the disclosure duty requires a separate analysis it does not resolve in the abstract.

That mirrors a pattern PPC Land has tracked across Article 50 coverage generally: obligations are distributed across providers and deployers, but production chains often involve several organizations performing different roles on the same asset, and the regulation does not always make clear which obligation controls when responsibilities overlap. VIA Nederland closed its guidance by inviting agencies with remaining questions to contact the association directly, treating the scenario-based framework as a starting point for individual assessment rather than a substitute for case-specific legal review.

Timeline

Summary

Who: VIA Nederland, the Dutch industry association for advertising, formally Verenigd In Advertising, publishing guidance for creative agencies operating in the Netherlands and the wider European Union.

What: A practical breakdown of when the AI Act's Article 50 transparency obligations require agencies to disclose AI use, organized around four scenarios: interactive AI such as chatbots and brand assistants, AI-generated image, audio and video content tested against a deepfake definition, AI-generated text on matters of general interest, and additional duties for agencies that develop their own generative AI systems.

When: Published September 5, 2026, more than a month after Article 50 obligations became legally applicable across the European Union on August 2, 2026.

Where: Applies to creative agencies and advertisers operating in or serving the European Union and European Economic Area, with VIA Nederland's guidance specifically addressed to the Dutch advertising industry.

Why: Article 50 does not require labeling every piece of AI-assisted content, only content meeting specific legal thresholds tied to an organization's role as provider or deployer, the type of content produced, and whether that content could deceive a viewer about its authenticity. Getting that distinction wrong carries two opposite costs: unnecessary labeling that research shows can reduce consumer trust and click-through, or missed disclosure that carries penalties reaching 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.