The AI Act labelling obligation since 2 August 2026: what you must label, and what is explicitly exempt

§01Since when does the AI Act labelling obligation apply?

Since 2 August 2026. The transparency obligations apply two years after the AI Act entered into force, and they apply to every affected AI system on the Union market, regardless of when it was placed there. There is exactly one exception.

The only transitional period, and who it does not cover

That deadline is not in Article 50 but in Article 111(4) of the AI Act, inserted by amending Regulation (EU) 2026/1744 of 8 July 2026. It says that providers of systems which generate synthetic content and were placed on the market before 2 August 2026 “shall take the necessary measures to comply with Article 50(2) by 2 December 2026”. Two restrictions, both from paragraph 153 of the guidelines: the deadline covers providers only, not you as a user. And systems that are partly interactive and partly generative benefit only for the machine-readable marking; the notice that you are talking to an AI was due on 2 August.

Does the obligation apply retroactively?

No. Content generated before 2 August 2026 does not have to be marked or labelled after the fact. One exception is in paragraph 154: a text generated earlier but published on or after that date must be labelled. The trigger is publication, not generation.

§02Who has to label, the provider or the deployer?

The AI Act distinguishes two roles, and almost every duty that reaches an ordinary company attaches to the second one. Providers develop AI systems and put them on the market. A deployer is whoever uses an AI system under their own responsibility.

Hiring an agency does not automatically make you a deployer. A company that merely commissions an advertising agency, without deciding on the use of AI and controlling it, is not a deployer; the duty then falls on the agency. Anyone who specifies that AI is to be used does decide and control.

The machine-readable marking does not relieve you. Under paragraph 117, deployers expressly cannot rely on the marking embedded by the provider. Metadata and watermarks are not perceptible to people; a visible or audible label is needed.

Only purely private, non-professional use is exempt. Send an AI image into the family chat and you are not a deployer. Put it on the company website and you are.

§03The four obligations of Article 50 at a glance

Under paragraph 5, the information must be provided clearly and distinguishably at the latest at the time of the first interaction or exposure, and it must meet the applicable accessibility requirements. A notice that a screen reader does not read out does not satisfy the duty. What that means in practice is in our article on the European Accessibility Act and accessible websites.

§04Label or not? 16 cases from everyday practice

Every row carries its source. “Para.” refers to the paragraph numbers of the Commission guidelines of 20.07.2026, section numbers to their structure.

The pattern behind it: the duty to label begins where the perception of reality is affected. Language polish, formatting and advertising copy fall outside. A face, a voice, a product that looks better than it is, that falls inside.

§05What is a deepfake, and what is expressly not one?

Under Article 3(60) of the AI Act, a deepfake is AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Both features have to come together: resemblance to something real and the capacity to deceive about its authenticity.

Paragraph 116 adds a rule of thumb for advertising: background extensions, aesthetic background replacement, arranging existing products and rescaling usually have little influence on how authentic an image appears. For journalistic images the opposite is true.

The contradiction between February and July 2026

Until the summer, the leading German-language practice guide on this subject was the guide published on 4 February 2026 by the Wettbewerbszentrale, the German association against unfair competition. It reads the definition of a deepfake noticeably more widely.

This is not an error on either side. It is a timestamp. The Wettbewerbszentrale wrote five months before the guidelines and says so itself: it sets out “how we currently understand the relevant provisions”, and over time court rulings would show whether the courts agree.

In practice that means: follow the wider reading and you label more than necessary. Legally that does no harm. It does cost effect, because a notice that stands everywhere is read nowhere. Follow the guidelines and you label precisely. The guidelines are the more recent and the more specific document, but they are non-binding.

§06When AI text has to be labelled, and when it does not

The text obligation in Article 50(4), second subparagraph, is considerably narrower than its reputation suggests. It applies only where three conditions come together: the text is published, it is meant to inform the public, and it concerns a matter of public interest.

Published means, under paragraph 131, accessible to an indeterminate and sufficiently large number of unconnected readers. Private professional correspondence and texts internal to an organisation, with publications on the internal company network named expressly, are not published.

Public interest means subjects that are relevant to society and worthy of public debate or scrutiny: politics, public administration, justice, fundamental rights, public security, health, environmental protection, consumer safety, and economic, financial, scientific and cultural developments.

The exception that settles the rest: even where all three conditions are met, the duty falls away if two things come together. The text has been reviewed for content by a natural person with the relevant expertise, and a natural or legal person carries editorial responsibility. Both together, not one of them.

For a company blog that means: if you genuinely edit your articles and assign responsibility by name, you need no AI notice. If you publish unreviewed, you need one, and you have a different problem anyway.

§07Do I have to label my chatbot as AI?

Yes. This is the clearest obligation in the whole article and at the same time the one most often overlooked. The guidelines name chatbots in online platforms and helpdesk tools expressly as an example where the obviousness exception does not apply.

The reasoning in paragraph 45 is plain: the exception is to be read narrowly, because it takes away people's right to information. General knowledge that chatbots exist does not mean that people recognise them in the specific interaction. The exception applies only where practically no doubt remains about the nature of the interaction.

It plainly applies to internal assistants for trained staff, to code assistants for professional developers and to AI characters in single-player video games. It does not apply to realistic avatars, to human-like voices, and it does not apply to the support chatbot on a public website.

Implementation: a notice on the chat window or in the first message, before the first user input. Not in the terms and conditions, not in the imprint. How that fits cleanly into an existing interface is part of our web design work.

§08How do I label correctly? Form, place, language

The label has to be perceptible to people without them opening a tool or taking a separate action, so visible or audible; metadata alone is not enough. And it has to appear at the latest on first exposure.

Place: at the image or video, not in the imprint. A note in the picture credits at the foot of the page is not “clear and distinguishable” when the image sits at the top.

Language: the Wettbewerbszentrale points out that higher regional courts have objected to English labels in influencer marketing (OLG Celle, judgment of 08.06.2017, 13 U 53/17; Kammergericht, decision of 11.10.2017, 5 W 221/17). If you address the German market, label in German: “KI-generiert” rather than “AI-generated”.

Accessibility: Article 50(5), second sentence, requires the information to meet the applicable accessibility requirements. A notice that sits as text inside a graphic, or exists only as a colour marker, does not satisfy the duty.

Are the EU icons mandatory?

No, and the Commission says so on its own page in one sentence: use of these EU icons is optional, the labelling duties under Article 50 of the AI Act are not. The symbols are an offer, the duty behind them is not.

Three symbols are offered. They are available as SVG and PNG, in black and white, each solid and semi-transparent, and anyone may use them freely, expressly without any attribution requirement towards the Commission or the AI Office.

There are no separate symbols for image, video, audio and text; the same three apply to every format. If you use them, the Commission's specifications require you to place them so that they are clearly perceptible at the latest on first exposure, are not hidden by overlays, survive sharing and downloading, and carry the usual accessibility means: alternative text and, where needed, an ARIA label.

The code of practice on transparency of AI-generated content is voluntary as well. The Commission published it on 10 June 2026 and assessed it in July as adequate within the meaning of Article 50(7); by 31 July 2026 around 190 organisations had signed it. Signatories are bound by its placement requirements, and anyone who has not signed may still use the icons.

For non-signatories the guidelines name two practical consequences: authorities have less insight and are expected to make more requests for information and access. And compliance with a code assessed as adequate can be taken into account as a mitigating factor when a fine is set.

§09What happens if you do not label?

Breaches of the transparency obligations in Article 50 can be fined up to 15 million euros or up to 3 per cent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises including start-ups, the lower of the two figures applies.

In Germany, the KI-MIG governs competence: the Act on Market Surveillance and Innovation Promotion for Artificial Intelligence of 22 July 2026, in force since 29 July 2026. Under section 2(1), the Bundesnetzagentur, the Federal Network Agency, is competent “unless otherwise provided in this Act”. Otherwise is provided for harmonised products, for supervised financial activities (BaFin), for public bodies of the federal states and for media providers. For the case groups of Article 50 the Act makes no special provision, so the Bundesnetzagentur remains competent. The authorities act on their own initiative or on a complaint; under Article 85 of the AI Act, anyone with indications of a breach may complain.

Next to that stands a second route. The Wettbewerbszentrale takes the view that breaches of the AI Act may at the same time be unfair competition, so that competitors and associations can bring injunction claims; it is itself entitled to sue under section 8(3) no. 2 of the UWG, the German Act against Unfair Competition. There is no court decision on this. We report it as a legal opinion, not as the legal position.

Independently of that, one rule applies that is not new: if you promise AI, you have to deliver AI. Advertising a product as “AI-based” without an AI component being present is prohibited misleading advertising under the UWG. And labelling under Article 50 does not make content lawful: a labelled deepfake stays unlawful if it infringes trade mark, copyright or personality rights.

§10What we do as an agency ourselves

No chatbot on our site. There is none, so Article 50(1) does not apply to us. That is not a compliance measure but a design decision taken before the AI Act: if you have a question, you should get an answer from one of the two people who also do the work.

Images. Where we use AI in photo and video production, we check every motif individually against the case groups in the table above and record the decision. We do not label retouching that falls under paragraph 92; anything that shows a product or a person differently from reality, we label visibly at the image.

This article itself. The illustrations above are AI-generated. They show no real person, no real product, no specific place and no actual event, so under Article 3(60) of the AI Act and Section 6.1.1 of the guidelines they are not deepfakes and carry no labelling duty. We write it down anyway, because the reasoning is worth more than the label. On a client project with a real product in the picture, the same check would come out differently.

Text. Our articles are reviewed for subject matter before publication, and editorial responsibility rests by name with the author. Her name stands above the article and once more in the box at the end of the text, not in the imprint. That brings the exception from paragraph 133 into play.

What we change for clients. Since 2 August, every project involving image, text or video production carries an extra line in the handover document: which content was generated with AI, which case group applies, and who the deployer is within the meaning of the regulation. That last question decides whether the duty sits with the client or with us, and it belongs in the contract, not in a later conversation.