The EU AI Act's Article 50 Labeling Duty for AI-Generated Marketing Content (Beyond Deepfakes)

Article 50 is several duties, not one

The most common misconception about the EU AI Act (Regulation (EU) 2024/1689) and marketing content is that "AI-generated content must be labelled" is a single, unified rule. It isn't. Article 50 contains several separate duties aimed at different parties — and for ordinary AI marketing content, two of them matter most:

• Article 50(2) requires the provider of an AI system (the company that builds the tool — an image or video generator such as Midjourney, DALL-E or Sora) to ensure its outputs are marked in a machine-readable format and technically detectable as artificially generated. This typically happens automatically, as an invisible watermark or embedded metadata, without the user doing anything.
• Article 50(4) requires the deployer — the brand, agency or influencer actually using the tool for a commercial purpose — to give the public a visible or audible disclosure when the output is a "deep fake": AI-generated or manipulated image, audio or video content that resembles an existing person, object, place, entity or event, and would falsely appear to a person to be authentic (Article 3(60)).

The distinction matters in practice: the provider's automatic marking happens regardless of what you make. Your own disclosure duty to your audience only kicks in if what you've made actually meets the deepfake definition — and most ordinary AI marketing visuals don't.

When is it "just" an AI image — and when does it become a deepfake?

The definition in Article 3(60) is narrower than most people assume. It requires the content to resemble an existing person, object, place, entity or event — not merely that the content is artificially generated. An AI-generated product mockup, a stylised background for an Instagram ad, or an invented, generic landscape is synthetic content, but it is not a deepfake, because it doesn't purport to depict anything specific and real.

That flips the moment the subject starts to resemble something or someone real: an AI-generated "photo" of your CEO that was never taken; an AI-generated clip that makes a model appear to be visiting a specific, named holiday resort she has never been to; or — the most obvious case in influencer marketing — an AI-generated video using a real creator's face or voice. If that's your situation, see our articles on unauthorized AI deepfakes of an influencer and AI voice cloning and dubbing of creator content, both of which cover exactly that situation.

Ordinary AI marketing content vs. a deepfake: what's actually required of you?

AI-generated text: the exemption that rarely rescues ordinary marketing copy

Article 50(4) also contains a rule about AI-generated text: whoever publishes AI-generated or AI-manipulated text with the purpose of informing the public on matters of public interest must disclose that the text is artificially generated. There is an exemption: the disclosure duty falls away if the text has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for it.

This is where marketers tend to misread the rule in both directions. First, the text duty itself only applies to text that informs the public on matters of public interest — ordinary sales copy, product descriptions or an Instagram caption about a sale typically fall outside that scope entirely, because they aren't intended to inform on a matter of public concern. Second, the human-editorial-review exemption is precisely that — an exemption from the text rule only. It doesn't extend to images, audio or video, and a marketer who has "reviewed" an AI-generated product photo does not thereby exempt that image from the general marking logic in Articles 50(2) and 50(4).

In practice: go ahead and use AI to draft ad copy, product descriptions or social captions — that typically doesn't trigger Article 50's text rule at all. Keep thinking carefully, on the other hand, whenever AI generates images or video that come close to resembling a real person, place or event.

The Code of Practice: the voluntary guidance behind the rules

On 10 June 2026 the European Commission published the final version of its Code of Practice on Transparency of AI-generated Content — a voluntary code that spells out concretely how providers and deployers can meet Article 50, including technical approaches to watermarking and metadata and practical requirements for what a visible disclosure should look like (among them a minimum 4.5:1 contrast ratio for text/image disclosures). The code is not binding law in itself, but the Commission and the AI Board have confirmed it as a recognised tool for demonstrating compliance with Article 50 — in practice, the reference point a regulator is likely to look for.

Penalties for breaching Article 50

Breaching the transparency duties in Article 50 is sanctioned under Article 99 with fines of up to €15 million or 3% of the company's total worldwide annual turnover for the preceding financial year, whichever is higher. That's a lower tier than the up-to-€35-million/7% penalty reserved for the eight specifically prohibited AI practices in Article 5, but still substantial.

The maths below is invented, for illustration only. Picture two companies that both breach Article 50: Company A has €30 million in global annual turnover. 3% of that is €900,000 — well below the flat €15 million ceiling, so the €15 million figure is what actually caps the fine in this case, not the percentage. Company B has €1 billion in global annual turnover. 3% of that is €30 million — above the €15 million floor, so here the percentage figure becomes the real, higher ceiling. The point: the larger the company, the more is genuinely at stake, because the fine is "whichever of the two numbers is higher", not a flat amount.

Decision framework: do you need to label this AI asset?

ASK FIRST: does the subject resemble a specific, identifiable existing person, a real place or a real event closely enough that an ordinary viewer could believe it's authentic? If no → you're dealing with ordinary synthetic content. No disclosure duty under Art. 50(4) — only the provider's technical marking applies, and that requires no action from you.

IF yes → the content is a deepfake in the legal sense. Add a visible or audible disclosure at first exposure, repeat it periodically for video/audio, and make sure it's easy to perceive without any special tools (see the contrast requirement above).

IF it's AI-generated text → ask whether the purpose is to inform the public on a matter of public interest. If it's ordinary sales or product copy, the text rule typically doesn't apply at all. If it's editorial content on a matter of public interest, and a named person holds editorial responsibility after human review, you can rely on the exemption.

IF you're unsure → add the visible disclosure anyway. The cost of an extra line of text or a small icon is low compared with the risk of a breach that reaches 3% of your global turnover.

How this differs from the other AI articles in the Academy

This article covers the broad, everyday case: a marketer using a generative AI tool to produce an image, video or piece of text with no specific, real person involved. That's a different situation from the two closely related Academy articles: AI voice cloning and dubbing of creator content covers a brand that, with consent, clones a real, named creator's voice, and unauthorized AI deepfakes of an influencer covers a third party making an unauthorized deepfake of a real influencer without consent. Both of those articles meet Article 50(4)'s deepfake rule by definition. This article covers all the AI marketing content that doesn't — and why the rule is still worth knowing, even when it doesn't apply to you.

Make Influence's operational perspective

The mistake we see most often from brands runs in two opposite directions. Some assume all AI-generated marketing content now needs a visible label "just to be safe" — and end up putting unnecessary AI tags on ordinary product images, which in our experience can create needless doubt in the consumer's mind about how much of the content is genuine at all. Others go the other way and assume that because "it's just an AI image", the rules simply don't apply to them — and miss that the exact cases where an AI image resembles a specific, real person or place are precisely where the disclosure duty bites hardest. Our recommendation: actively run the decision framework above every time you release AI-generated visual material, rather than adopting a blanket policy in either direction.

FAQ

Do we need to label all our AI-generated ad images as artificial?

No, not automatically. Only if the image meets the deepfake definition — that is, if it resembles a specific, identifiable existing person, place or event closely enough to falsely appear authentic. A generic AI product image or an AI-generated background typically doesn't trigger your disclosure duty.

Does Article 50 apply to AI-written ad copy and product descriptions?

Rarely in practice. The text rule in Art. 50(4) only applies to text intended to inform the public on matters of public interest — ordinary sales copy typically falls outside that scope.

Who is responsible if we use an AI image tool that has no built-in marking of its own?

The technical marking duty (Art. 50(2)) sits with the provider of the tool, not with you as the user. Your own responsibility is the disclosure duty toward your audience (Art. 50(4)) if the content is a deepfake — and you can't pass that one to the vendor.

Is the AI labeling Code of Practice a legally binding law?

No. The Code of Practice on Transparency of AI-generated Content, finalised by the Commission on 10 June 2026, is voluntary — but the Commission has confirmed it as a recognised tool for demonstrating Article 50 compliance.

What is the penalty for breaching Article 50?

Fines of up to €15 million or 3% of the company's global annual turnover, whichever is higher, under Article 99.

Does Article 50's machine-readable marking mean the same thing as C2PA?

Not necessarily — Article 50(2) requires machine-readable marking but doesn't mandate a specific technical standard. C2PA's Content Credentials is one way providers meet that requirement in practice, and it's also used by platforms independently of the AI Act. See Content Credentials (C2PA) explained for how the technical standard works, who's behind it, and its real-world limits.