Law · AI · 6 min read

Labelling AI content What Article 50 requires and what it does not

The transparency obligations of the EU AI Act have applied since 2 August 2026. Since then two sentences have been circulating, and both fall short: „everything has to be labelled now“ and „this only concerns the big AI providers“. What Article 50 actually says is narrower in several places and more uncomfortable in one.

Simon Förstemann Growth Strategist · Lake Constance Region August 2026 As of 13 August 2026

Key Takeaways

· Article 50 distinguishes between providers (who develop and offer the AI system) and deployers (who use it professionally). If you do marketing with AI, you are a deployer.
· The machine-readable marking of synthetic content is an obligation of providers under Article 50(2), not yours.
· As a deployer, Article 50(4) gives you a disclosure duty in two cases: deep fakes, and AI-generated text published to inform the public on matters of public interest.
· Plain marketing and product copy is not named there. And where a human reviews the content and holds editorial responsibility, the text duty falls away explicitly.
· Fines run to 15 million euro or 3 % of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, Article 99(6) applies the lower of the two.

Orientation, not legal advice This text reflects my research as of 13 August 2026 and does not replace a lawyer reviewing your specific case. The interpretation of Article 50 is young in several places; supervisory practice and case law are largely still to come. If you need a binding answer, get it from a qualified lawyer.

Who is actually addressed?

The most common error in the current debate is that providers and deployers get thrown together. The regulation separates them clearly.

A provider develops an AI system, or has one developed, and offers it under its own name. Those are the model houses and tool vendors. A deployer uses an AI system in a professional capacity. That is you, the moment you generate text, images or video for your business with an AI tool. An agency producing for you is a deployer too.

This distinction decides which duties apply to you at all. The technically demanding one, machine-readable marking, is not yours.

The four duties, sorted by who is addressed

Paragraph 1, providers: make it recognisable that someone is talking to an AI. Applies to systems interacting directly with people, chatbots for example. Falls away where it is obvious anyway to a reasonably well-informed user.

Paragraph 2, providers: mark synthetic content in a machine-readable format. Covers audio, image, video and text from generative systems. Systems performing an assistive function for standard editing, or not substantially altering the input data, are exempt. This is why the large providers embed watermarks. The duty does not sit with you as a user.

Paragraph 4, deployers: disclose deep fakes. This is where it gets concrete for companies. The regulation defines a deep fake narrowly, more on that below.

Paragraph 4, deployers: disclose AI-generated text, but only in one case. Namely where the text is published to inform the public on matters of public interest. And even then the duty falls away where the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for it.

Paragraph 5 adds a question of form: where disclosure is required, it has to happen in a clear and distinguishable manner, at the latest at the time of first exposure. A note in your legal notice does not cover it.

What does this mean for ordinary marketing?

If you write a product description, a newsletter or an ad with AI assistance, and you read and stand behind it before publishing, the text duty in paragraph 4 does not bite on the wording. First, that is not a contribution informing the public on matters of public interest. Second, it is exactly the human review the provision explicitly carves out.

The practical core The provision does not target the use of AI. It targets deception about where a piece of content came from. Anyone using AI as a tool and taking editorial responsibility for what they publish sits outside the text disclosure duty in the normal case.

An editorial piece on a matter of public interest is a different thing. If you use AI for that and nobody holds editorial responsibility, the duty applies.

Where it really gets tight: the deep-fake definition

Under the regulation, a deep fake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events, and that would falsely appear authentic to a viewer. Two elements have to come together: a reference to something that really exists, and the appearance of authenticity.

That is where it gets expensive for companies. An AI image that looks like a real photo from your premises, an AI face appearing as a satisfied customer, an AI-generated before-and-after: those are the cases that come into view on the wording.

Conversely, purely symbolic imagery with no real-world template, no recognisable person, no identifiable place, no specific event, lacks the reference to something existing on this reading. Whether supervisory authorities will classify photorealistic symbolic images the same way is not settled. If you need certainty here, get it in writing from a lawyer or label pre-emptively. Labelling does no legal harm. It only costs some effect.

What applies regardless of the AI Act?

I consider this the most important point, because it is older and bites more often: presenting an AI image as a genuine customer photo was already a misleading commercial practice under unfair competition law long before the AI Act. Actionable by competitors, independent of any supervisory authority.

In practice: the AI Act sharpens a risk that already existed. Anyone following the simple rule of never presenting AI content as something it is not has dealt with most of both topics.

An approach that holds

Sort your content by its reference to reality. Anything showing, or purporting to show, real people, real places or real events goes on the list for review. Purely symbolic imagery is the less critical case.

Settle who is responsible for text. A named person holding editorial responsibility is the simplest safeguard, and it is good practice anyway.

Separate illustration from evidence. Testimonials, references, before-and-after: whatever serves as proof should be real. Whatever serves as illustration may be generated, and should never look like proof.

When in doubt, label or get legal advice. The interpretation is young. A visible label costs you a little effect. Getting it wrong costs more.

On the level of fines Public discussion usually leads with 15 million euro. That is correct as a ceiling for breaches of Article 50, but it applies alongside 3 % of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, Article 99(6) provides that the lower of the two applies. For an SME the realistic ceiling is therefore turnover-dependent and well below the headline.

Frequently Asked Questions on Labelling AI Content

Do I have to label every AI image on my website?

The regulation does not require that across the board. The disclosure duty for deployers attaches to the notion of a deep fake, which presupposes a reference to real existing persons, places or events. Purely symbolic imagery with no real-world template lacks that reference on the wording. For photorealistic symbolic images the question is not settled, so a legal opinion or pre-emptive labelling is the safe route if you need certainty.

Do I have to label AI-written marketing copy?

On the wording of Article 50(4), the text duty concerns publications intended to inform the public on matters of public interest. Plain marketing and product copy is not named there. In addition, the duty falls away where the content underwent human review or editorial control and someone holds editorial responsibility for it.

Do I have to add the watermark myself?

No. Machine-readable marking of synthetic content is addressed to the providers of AI systems under Article 50(2), not to you as a user. Your duties as a deployer sit in paragraph 4.

What happens if I use an AI image as a genuine customer photo?

That is the clearest problem case, and it has two sides. Under the AI Act a disclosure duty comes into view, because there is a reference to a person appearing real. Independently of that, it is a misleading commercial practice under unfair competition law and therefore actionable by competitors. That risk existed before the AI Act.

Since when does this apply?

The transparency obligations in Article 50 have applied since 2 August 2026. Supervisory practice and case law are largely still to come, and the interpretation of individual terms may yet develop.

The reminder, once more This piece is orientation as of 13 August 2026, not legal advice. It describes the wording of the regulation and names where the interpretation is open. For your specific plans it does not replace a lawyer.

Bereit für den nächsten Schritt?

Book an intro call

30 Minuten, in denen wir beide prüfen, ob eine Zusammenarbeit passt.

Schedule a call →

Case Study +74 % revenue in 16 months. Same budget, 12 % fewer visitors. How I did it →

About the author

Simon Förstemann

Simon Förstemann weiß, wie es ist, an deiner Stelle zu stehen: 14 Jahre Erfahrung, 6 eigene Gründungen, 3 Exits, Red Dot, German Design Award und German Brand Award 2026. Dieses Knowledge bringt er 1:1 in dein Unternehmen, damit dein Marketing verkauft statt nur zu beschäftigen. Er berät nicht nur, er unterstützt dich auch operativ bei der Umsetzung. Keine Agentur, keine Workshops, die nichts bringen.

LinkedIn →