EU AI Act Article 50 for Agencies: 4 Triggers That Decide When You Label AI Content
The direct answer
Since 2 August 2026, Article 50 of the EU AI Act is enforceable law across the EU. But it does not force your agency to label every AI-assisted asset: the obligation is role-based and use-case-based, not tool-based. Most agencies are deployers, and deployer duties fire on four kinds of work — chatbots and brand assistants, deepfake image/audio/video, public-interest text, and custom AI tools built on a model (where you can tip into provider territory). Non-compliance carries fines up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher.
In August 2026, VIA Nederland — the Dutch advertising trade body — published guidance for creative agencies on when EU AI Act transparency rules force AI disclosure. Its answer is four triggers covering nearly every way an agency touches generative AI. Here is what each means, with agency examples and the disclosure each requires.
Article 50 is live: what changed on 2 August 2026
The AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024; Article 50 became applicable EU-wide on 2 August 2026. The European Commission's final Article 50 guidelines (C(2026) 5054 final) were published 20 July 2026, alongside confirmation that the Code of Practice on Transparency of AI-generated Content (final text 10 June 2026) is adequate. Three timing details matter:
- Deployer duties were live from day one. The 2 December 2026 grandfathering date covers only providers' machine-readable marking of systems placed on the market before 2 August 2026.
- No retroactive labelling. Content generated before 2 August 2026 needs no retroactive label — except AI public-interest text first published on or after that date.
- Fines are real. Article 99(4) sets fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher, for Article 50 breaches.
Know your role first: deployer or provider
A provider develops an AI system and places it on the market under its own brand. A deployer uses an AI system for professional purposes. VIA's starting point:
"Most creative agencies will be deployers. They generally do not build an AI system themselves but use existing systems such as ChatGPT, Midjourney or Adobe Firefly." — VIA Nederland, 31 August 2026 (translated from the Dutch original)
The split matters because the duties differ: providers must make synthetic output detectable in a machine-readable format (Article 50(2)); deployers owe the visible, human-perceivable labels for deepfakes and public-interest text (Article 50(4)). A deployer cannot satisfy a visible label by pointing at the provider's invisible marking. Build your own generative tool on a model and each application must be assessed individually — you may be a provider for that tool.
The 4 triggers for AI disclosure
Trigger 1 — Chatbots and brand assistants: say it is AI at first contact
What it is. When an AI system talks directly to people — a brand chatbot, an AI assistant, an AI persona — users must know they are talking to AI (Article 50(1)), unless it is obvious from the context. A consumer chatbot styled to feel human rarely gets the "obvious" pass.
Agency example. You launch a hotel-client concierge chatbot trained on the brand voice that greets guests by first name. Disclosure belongs where the conversation starts.
Required action. A clear AI notice at first contact — short text for chat, spoken for voice. A mention buried in terms and conditions does not count. If the client uses an existing assistant rather than one you built, map which party in the chain owes the disclosure.
Trigger 2 — AI images, audio and video: the deepfake test
What it is. Article 3(60) defines a deepfake as AI-generated or manipulated image, audio or video "that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful." Plainly: could a viewer be deceived about authenticity? The AI person does not need to match a known real individual. Clearly fictional scenes and minor post-production are not deepfakes.
Agency example. A commercial with photoreal AI talent in a realistic setting, or a synthetic voice resembling a known presenter, needs a label. A clearly fictional animated campaign does not — the Commission's own example treats AI video of cartoon mice arguing over cheese in an ad as not a deepfake.
Required action. Disclose at first exposure in a clear, distinguishable, human-perceivable way. Metadata or an invisible watermark alone is not enough. Plan the visible label into the creative; Commission icons are optional.
"For creative advertising messages, this exception will unfortunately not apply." — VIA Nederland on the artistic-content carve-out, 31 August 2026 (translated from the Dutch original)
The carve-out trap: Article 50(4) allows less prominent disclosure for "evidently artistic, creative, satirical, fictional" work — but not where the commercial message dominates.
Trigger 3 — AI-drafted text on matters of public interest
What it is. Deployers owe a text label only when AI generates or manipulates text published to inform the public on matters of public interest — politics, public health, public safety, fundamental rights, public debate (Article 50(4)). For agencies this usually means societal and public-awareness campaigns, not routine commercial copy.
Agency example. You draft a public-health informational article for a healthcare client partly with AI. The client either labels it as AI-generated or documents substantive human editorial control.
Required action. Label the text unless it "has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication." A spellcheck is not enough — record who substantively reviewed it and when.
What does not trigger it. Ordinary commercial copy, blogs and social posts sit outside the deployer text duty. As PPC Land put it: "Not every commercial made with Midjourney needs a label, and not every script drafted with ChatGPT does either."
Trigger 4 — Custom AI tools on a model: provider territory
What it is. An agency that builds its own AI system or an application layered on an existing model may qualify as a provider for that tool — the least common but most consequential scenario.
Agency example. You wrap an API model into a branded "campaign copy generator" sold to clients. That product may make you a provider; using the same API directly to draft one client's copy keeps you a deployer.
Required action. Providers must ensure synthetic output is "marked in a machine-readable format and detectable as artificially generated or manipulated" (Article 50(2)). The Code of Practice calls for at least two marking layers for most content — digitally signed, time-stamped metadata plus an imperceptible watermark, with text over 200 tokens also watermarked. Marking is not required where AI only assists standard editing. Adherence to the Code is voluntary, but signatories — Google, Meta and Anthropic among ~190 — can rely on it to show compliance.
The four-trigger checklist
| # | Trigger | Typical agency work | Label needed? | Required action |
|---|---|---|---|---|
| 1 | Chatbot / brand assistant / AI persona | Client concierge chatbot, voice assistant, branded persona | Yes, unless "obvious" it is AI | Clear AI notice at first contact — not buried in terms and conditions |
| 2 | AI image / audio / video (deepfake test) | Photoreal AI talent, synthetic voice, realistic AI scenes | Yes if a viewer could be deceived about authenticity | Visible AI label at first exposure; metadata or invisible watermark alone is not enough |
| 3 | AI-drafted text on public-interest matters | Public-health, political, public-awareness campaigns | Yes for public-interest text | Label as AI-generated OR document substantive human review with named editorial responsibility |
| 4 | Custom AI tool built on an existing model | Branded generator or agent product sold to clients | Provider duty for that tool | Machine-readable, detectable marking of synthetic outputs (metadata + watermark per Code of Practice) |
Routine commercial copy, blogs and social posts that fit no trigger generally need no deployer label.
FAQ: labeling AI content under the EU AI Act
Do we have to label every AI-generated ad or ChatGPT draft?
No. Article 50 duties are role-based and use-case-based, not tool-based: not every commercial made with Midjourney needs a label, and not every script drafted with ChatGPT does either. Only specific uses can trigger a duty.
Who is responsible for disclosure — our agency or the AI tool provider?
Most agencies are deployers who use tools like ChatGPT, Midjourney or Adobe Firefly, so they carry the visible labelling duties for deepfakes and public-interest text. Providers carry machine-readable marking duties. A deployer cannot rely on the provider's invisible watermark to satisfy a visible label. Building a custom tool on a model can make you a provider for that tool.
What counts as text on "matters of public interest"?
Text published to inform the public on politics, public health, public safety, fundamental rights or other matters of public debate. For agencies this usually means societal and public-awareness campaigns, not routine commercial copy.
What makes an AI image or video a deepfake under Article 50?
The test is whether AI-generated or manipulated content resembles existing persons, objects, places, entities or events and could falsely appear authentic. The AI person need not match a known real individual. Clearly fictional content is not a deepfake, and the artistic carve-out does not rescue content where the commercial message dominates.
What about content we made before 2 August 2026?
No retroactive labelling is required, with one exception: AI text on public-interest matters first published on or after 2 August 2026 does need a label. The 2 December 2026 grandfathering applies only to providers' machine-readable marking of systems placed on the market before 2 August 2026.
What are the penalties for getting Article 50 wrong?
Fines of up to EUR 15 million or, for undertakings, up to 3% of total worldwide annual turnover — whichever is higher — for breaches of the Article 50 transparency obligations, applied by EU member state authorities.
What your agency should do now
VIA's five-point routine makes transparency part of the creative process: determine your role; look at what the AI actually does; run the authenticity check on visuals and audio; document who reviews AI output and carries publication responsibility; and decide whether — and how — to disclose without disrupting the creative. When a trigger fires, build the disclosure into the deliverable and keep the evidence in the client file: the first-contact script, the label placement, the named editor. And when vetting an agency, ask whether it knows which triggers its client work hits — and can show you the audit trail. To run the four triggers as a scored client-work audit — with the deployer/provider role check, the responsibility-chain map and an evidence field — use the AI Content Disclosure Audit checklist on My Business AI Audit.
Get matched with a vetted AI agency that treats transparency and compliance as part of delivery.
Browse Vetted AI Agencies →Sources
- [1] PPC Land — "Dutch trade body maps 4 AI disclosure triggers for ad agencies" (September 5, 2026): ppc.land — 4 AI disclosure triggers for ad agencies
- [2] VIA Nederland — "AI-transparantie voor creatieve bureaus" (guidance dated August 31, 2026, Dutch original; quotes used here are working English translations): vianederland.nl — AI transparency for creative agencies
- [3] European Commission — Guidelines on transparency obligations for providers and deployers of AI systems (July 20, 2026): digital-strategy.ec.europa.eu — Article 50 guidelines
- [4] European Commission — Code of Practice on Transparency of AI-generated Content (final code text June 10, 2026; adequacy confirmed July 20, 2026): digital-strategy.ec.europa.eu — Code of Practice
- [5] European Commission — Guidelines on the implementation of the transparency obligations under Article 50, C(2026) 5054 final (July 20, 2026): Commission guidelines C(2026) 5054 final (PDF)
- [6] EUR-Lex — Regulation (EU) 2024/1689 (EU AI Act), Articles 3(60), 50, 99(4), 113: eur-lex.europa.eu — Regulation (EU) 2024/1689
Accuracy note: Verified 2026-09-05 against the research brief for this story (kanban t_983ca14a; grounded-citations verify passed, 6/6 sources evidence-backed). Article 50 became applicable EU-wide on 2 August 2026 (Article 113, Regulation (EU) 2024/1689); guidelines C(2026) 5054 final are dated 20 July 2026; final Code of Practice text is dated 10 June 2026 with adequacy confirmed 20 July 2026. Fines are described as maxima ("up to EUR 15M or 3%"), applied by member state authorities. VIA quotations are working translations from the Dutch original (31 August 2026); the PPC Land quotation is from its September 5, 2026 coverage. Informational only — not legal advice.