# AI Labelling Duties: What Actually Applies as of 2 August 2026

> Article 50 of the AI Act has applied since 2 August 2026. What companies actually have to label, what expressly falls outside, and what applies additionally in Germany, Austria, Switzerland and Denmark.

*Published: 2026-08-20*

*Source: https://vellmerk.ai/en/blog/ki-kennzeichnungspflicht-art50*

Article 50 of the AI Act has applied since 2 August 2026. Two narratives have been circulating ever since, and both are wrong.

The first says every AI-generated text must now be labelled, and anyone who fails to do so faces up to 15 million euros. The second says it is nothing to worry about because it was postponed to December anyway.

If you have spent the past weeks trying to extract an answer to the question that actually matters, namely **whether you personally have to do anything**, you will know the problem. Most write-ups describe the regulation but not your situation. And some circulate claims that simply do not hold.

This article does something different. It answers four questions in order: Does this even apply to you? Which of your content is covered? What applies additionally in your country? And what is honestly still unresolved?

All central statements are sourced, wherever possible from primary sources. Where only secondary sources were available, or where something is unresolved, we say so.

> **Please note**  This article is a free information offering and **not legal advice**. It does not replace an assessment of your individual case and makes no statement about which obligations apply to you. There is as yet no case law on Article 50, and central terms remain unresolved. Detailed notes appear at the end of this article.

## 1. Does this even apply to you?

This is the question almost everything turns on, and most write-ups skip it.

Article 50 distinguishes two roles. **Providers** are the makers of the AI systems, so OpenAI, Anthropic, Midjourney, Adobe, Canva. **Deployers** are those who use such systems professionally, which includes you when you use ChatGPT for a text or Firefly for an image.

The regulation distributes the duties very unevenly:

**Paragraph 1 addresses the provider.** The system must make clear that one is interacting with AI.

**Paragraph 2 addresses the provider.** Marking outputs in a machine-readable format, plus means by which artificial generation can be detected.

**Paragraph 3 addresses the deployer.** Informing those affected where emotion recognition or biometric categorisation is used.

**Paragraph 4 addresses the deployer.** Disclosing deep fakes, and disclosing AI text on matters of public interest.

**For a mid-sized company using off-the-shelf tools, paragraph 4 is therefore what usually remains.** Paragraph 3 is added only for those deploying emotion recognition or biometric categorisation, more on that below. The much-discussed machine-readable marking, meaning watermarks and metadata, is the tool makers' job. Deployers do not have to build it in themselves. It can be different where a deployer becomes a provider, more on that in the chatbot section.

Three subtleties that are regularly misjudged in practice:

**Freelancers and agencies change nothing.** A legal person remains the deployer even where it involves third parties.

**Merely commissioning an agency does not make you a deployer**, as long as you do not take decisions and exercise control over whether and how the agency uses AI.

**A dual role is possible**, for instance where you adapt a purchased system with your own data and run it under your own name.

## 2. Which of your content is covered?

Paragraph 4 covers exactly two cases: **deep fakes** and **text on matters of public interest**. Everything else falls outside.

### Image, audio and video: the deep fake test

The wording expressly covers image, audio **and** video content. Synthetic voices are therefore covered on an equal footing.

A deep fake exists where three things come together. The content resembles persons, places, objects, entities or events that exist, could plausibly exist, or could plausibly have existed. It falsely appears authentic. And it is AI-generated or AI-manipulated.

The European Commission's guidelines of 20 July 2026 make this concrete with examples. **The following, among others, are not deep fakes:**

an AI-generated sphinx **flying** over the Eiffel Tower, because it is evidently impossible

talking mice in a cheese advertisement

a **real product against an AI-generated background**, as long as the advertisement does not mislead about the product itself

a radio broadcast with purely technical audio adjustment

**Deep fakes, according to the guidelines and the German competition watchdog, include:**

a **fully AI-generated product image** in advertising or on packaging that can mislead as to the product's appearance, characteristics or use

an AI video featuring a prominent person in an advertising context

a realistic synthetic avatar of an executive

AI influencers and synthetic models

The distinction from the point above matters. A **real** product against an AI background is unproblematic, an **AI-generated** product image is not. For many companies this is the most frequent case in practice.

Two clarifications from the guidelines that are often missing:

**Photorealism alone is not sufficient.** What matters is the capacity to mislead. The assessment is objective, an intention to deceive is not required.

One practical note on this: the German competition watchdog Wettbewerbszentrale, a private enforcement body, reads the term **more broadly** in its guidance and lets an abstract resemblance suffice where the image looks realistic overall. There is no case law on the point. Anyone relying on the Commission's narrower standard may still face a complaint.

**Minor editing generally stays outside.** According to the guidelines, AI colour correction, lighting adjustments, extending or replacing a background for purely aesthetic reasons and rescaling product images generally have only a minor impact. Removing passers-by in the background is expressly named as well.

Whether an edit goes beyond that depends on whether it changes the **perceived authenticity** of the content in the specific context. The guidelines contain no fixed list of impermissible interventions.

**And one point that is easily missed.** Where a deep fake exists, the tool maker's invisible watermark is not enough. The guidelines say expressly that deployers **cannot** rely on the machine-readable marking under paragraph 2 for the disclosure required by paragraph 4. A humanly perceptible indication is needed.

> **A confusion that happens often, and which we deliberately keep apart here.** Lists circulate naming interventions such as face swaps, voice cloning or altered body shape as supposedly marking the deep fake threshold. Those examples come from a different context, namely the **machine-readable marking duty under paragraph 2**, which applies to the tool makers. For the question of whether a deep fake exists under paragraph 4, the standard set out above applies.

### Text: the all-clear that matters most

This is the most important point for most companies, and it is rarely stated clearly enough.

The text obligation applies only where **three** conditions come together. The text is published, it informs the public, and it concerns a **matter of public interest**.

As examples of public interest the guidelines name politics, democratic processes, public administration, justice, fundamental rights, public security, health, environmental protection and consumer safety, as well as economic, financial, scientific or cultural developments that may be the subject of public debate.

**Expressly outside the scope, according to the guidelines:**

AI-manipulated text forming part of a company's advertising or product descriptions

advisory text on questions of regulatory compliance for clients

chatbot answers that reach only the person prompting

fantasy novels

A product text written with AI support is therefore, as a rule, **not** subject to labelling.

> **But there is a qualifier that is missing almost everywhere.** The exclusion for advertising applies, according to the guidelines, only so long as it contains no claims relating **for example** to **health, consumer safety or sustainability**. The list is expressly not exhaustive in the original; other sensitive claims can equally tip the assessment.

### The editorial exception, and the trap inside it

Even for texts on matters of public interest the obligation falls away where two things come together. The content has undergone a process of **human review or editorial control**, **and** a natural or legal person holds editorial responsibility.

The guidelines set the bar higher than many assume. What is required is a deliberate examination of the substance, with fact-checking treated as a minimum requirement. Expressly **not** sufficient are spell-checking, the mere existence of an editorial policy, automated review processes or cursory approval.

Two further points decide success or failure in practice:

**The order matters.** A substantive AI intervention **after** the human review causes the exception to fall away. The human must stand at the end of the process, not in the middle.

**The responsible person must be findable.** The guidelines require that the identity and contact details of the editorially responsible person or entity are publicly and easily findable, for instance in the website's legal information. The usual imprint details alone do not say who holds editorial responsibility. Precisely that naming has to be added.

### Legacy content: no retroactivity, but take care when publishing later

Content created **before** 2 August 2026 does not have to be labelled retroactively, according to the guidelines. There is no retroactive effect.

One exception matters in practice. Texts created before that date but **published afterwards** are to be labelled. Anyone working through a stock of pre-produced pieces should keep that in mind.

### Chatbots: the exception almost never applies

Paragraph 1 requires that people recognise they are talking to a machine. The guidelines read the exception for cases where this is obvious **narrowly**. General awareness that chatbots exist is expressly not sufficient.

As a negative example the guidelines expressly name chatbots embedded in online platforms and support helpdesks. For a publicly accessible chatbot on a company website the exception therefore does not apply in practice.

**Not sufficient as an indication are:** a note only in the terms and conditions, a machine-readable marking, a blanket "this website uses AI", or a purely technical description.

Two points follow from the allocation of roles. Paragraph 1 formally addresses the **provider**, that is the maker of the chatbot. In practice this means the website operator should check whether the embedded chatbot actually displays the notice, and add it otherwise. One higher regional court has classified the use of a chatbot as the website operator's own commercial act under competition law. That decision is not yet final.

> **An unresolved point that affects many.** At what stage does the deployer of a purchased chatbot become a provider? The German competition watchdog lets appearing under one's own name suffice, the Commission additionally requires a modification of the system, for instance through one's own training data. This matters in practice, because the provider role also carries the machine-readable marking duty under paragraph 2.

### Art and satire: attenuates, does not exempt

A widespread misconception. Where the content forms part of an evidently artistic, creative, satirical or fictional work, the transparency obligations are limited to disclosing its existence **in an appropriate manner that does not hamper the display or enjoyment of the work**. Disclosure is still required.

For advertising this hardly helps. The guidelines exclude content whose character is recognisably exclusively informative or commercial. Where content combines several elements, the informative character is to prevail, with the consequence that the regular labelling duty applies.

And a clarification that often gets lost: **this exception applies only to deep fakes, not to text.** For texts on matters of public interest there is no art or satire exception, only the editorial one.

### And the point where labelling is not enough

If you deploy emotion recognition or biometric categorisation, check something other than the information duty first.

**Emotion recognition in the workplace and in educational institutions is in principle prohibited under Article 5 of the AI Act.** The regulation exempts uses introduced for medical or safety reasons. Also prohibited in principle is biometric categorisation inferring race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation; here too narrowly drawn exceptions exist.

The decisive point: **informing those affected does not make a prohibited use lawful.** The transparency duty in paragraph 3 presupposes that the use is permitted at all.

## 3. The screening questions in brief

If you want to apply this framework to your own situation, work through these questions in order. They do not replace a legal assessment, but they sort out the position.

Are you using off-the-shelf tools, or developing your own system? With off-the-shelf, unmodified tools one is generally the deployer, and paragraphs 1 and 2 then address the maker.

Is the use professional or on behalf of an organisation? Purely private, non-professional use by natural persons falls outside, which follows from the definition of deployer in Article 3 point 4 of the regulation.

Is it image, audio or video? Then first: does the content resemble something that exists or could plausibly exist?

Could the audience take it for a real recording or a real voice?

Was the AI substantially involved, or is this minor editing?

Is the content evidently artistic or satirical? That attenuates the form of disclosure, it does not remove it.

Is it text? Then first: is it published, meaning accessible to an indeterminate, fairly large readership?

Does it concern a matter of public interest? Advertising and product descriptions generally do not, unless they contain claims relating for example to health, consumer safety or sustainability.

Was there a substantive human review, and is a responsible person publicly named and easy to find?

Did the human stand at the end of the process, or did AI run over the text again afterwards?

Is it a chatbot? Then first: is it publicly accessible or only for trained specialists?

Is the notice visible when the chat starts, and not only in the terms and conditions?

## 4. Four countries, four realities

The labelling duty itself is identical across the EU, because it sits in a regulation. What differs is everything around it: competence, the route to sanctions, and the national layer. And the differences are larger than one would expect.

A figure that is almost always missing from the debate: for the frequently cited penalty range of 15 million euros or three per cent of worldwide annual turnover, the **higher** of the two amounts applies as a general rule. For **small and medium-sized enterprises including start-ups the regulation reverses this**: there the **lower** amount applies. For most readers of this article the large figure is therefore simply not the relevant one.

**Germany.** The Federal Network Agency is the competent authority. There is no separate national penalty provision for Article 50; sanctioning runs through the directly applicable regulation. The greatest practical risk comes from associations, not from competitors.

**Austria.** **No authority has been designated** and there is **no national penalty provision**. The greatest practical risk is an injunction action under competition law.

**Switzerland.** There is no general AI supervision and **no labelling duty in domestic law**. The greatest practical risk is the AI Act itself, which can reach in through the market-location principle.

**Denmark.** The Agency for Digital Government is the competent authority. There is **currently no** national penalty provision for Article 50. The greatest practical risk lies in marketing law.

### Germany

The AI market surveillance and innovation promotion act has been in force since 29 July 2026, promulgated in the Federal Law Gazette 2026 I No. 223. The Federal Network Agency is the central market surveillance authority; for media service providers using AI for journalistic or advertising purposes it is the respective state media authority.

Notably, the act contains **no separate penalty provision** for Article 50. It merely determines the competent authority for the directly applicable penalties of the regulation. No fines are imposed on authorities and public bodies. That does **not** release them from the obligations themselves: supervisory measures remain possible, and every affected person has a right to complain.

On the risk of warning letters, a point missing from almost every write-up: for **competitors**, reimbursement of costs for warning letters concerning breaches of information and labelling duties in electronic commerce is excluded by the wording of the statute. One higher regional court has restricted that exclusion for a different labelling duty where a misleading practice is present at the same time. Whether that transfers to AI labelling is open. For **associations** with standing to sue the exclusion does not apply in any case, and the German competition watchdog set up a dedicated complaints office on 28 July 2026 with a category for deficient labelling of AI content.

Another frequent misconception: the social bot labelling rule in the German interstate media treaty concerns automatically posting accounts in social networks that appear to be personal profiles. **Own websites, blogs and newsletters do not fall under it on the wording.**

### Austria

Here the finding is the most surprising. Austria has **to this day designated no market surveillance authority and enacted no implementing act**. The deadline for this expired on 2 August 2025, so it is now more than a year overdue. The AI service point at the regulatory authority is an information hub without supervisory or sanctioning powers; it says so plainly on its own FAQ page.

This does **not** mean nothing can happen in Austria. It means the risk lies elsewhere. Possible claimants include, besides competitors, the Economic Chamber, the trade union federation, the Federal Competition Authority, the consumer information association and the association against unfair competition. Whether a breach of Article 50 qualifies as a breach of law for that purpose has not been decided. The route through the prohibition of misleading practices is more robust.

Austria has no apparent statutory regime for reimbursing the costs of such letters, comparable to German law. For that reason alone the German narrative of a wave of warning letters does not transfer.

### Switzerland

Swiss law contains **no labelling duty for AI content**. The Federal Office of Justice states this expressly in its legal baseline analysis and names it at the same time as one of the gaps for ratifying the Council of Europe convention.

Anyone concluding from this that they are not affected in any case is jumping to conclusions, however. The European Commission's guidelines expressly name **publishing on the globally accessible internet** as a connecting factor for deployers from third countries. A Swiss company that itself foresees the dissemination of its content in the EU, for instance by publishing on the freely accessible internet, can therefore fall under Article 50. Covered in that scenario are paragraphs 3 and 4. Not covered are distribution channels that are unforeseeable and outside your control.

Alongside this, technology-neutral norms apply: competition law for misleading statements, the protection of personality under the civil code where identifiable persons are involved, and the criminal offence of identity misuse, in force since 1 September 2023. Swiss criminal law has no deep-fake offence.

The Federal Council is working on a proposal, but it is unlikely to enter into force any time soon.

### Denmark

Denmark was the first EU member state with a national accompanying act, in force since 2 August 2025. Its penalty provision does **not** cover Article 50, however. A bill that would have changed this was introduced in February 2026 and lapsed with the close of the parliamentary session.

A structural difference hardly anyone mentions: in Denmark fines are imposed under **criminal** law, as a rule through a fixed-penalty notice that presupposes an admission of guilt. That is a different mechanism from the administrative procedure in Germany.

More relevant in practice at present is marketing law, which requires the commercial intent of any commercial practice to be disclosed clearly, and obliges businesses to document the accuracy of factual statements.

## 5. Fact check: three claims that do not hold

During our research we came across three claims circulating in the market, in part on law firm websites, that do not hold up against the primary sources. We deliberately name no names; the point is the substance.

**"The labelling duty was postponed to December 2026."** Wrong. The Digital Omnibus amended the AI Act in around forty places and postponed high-risk deadlines. **The obligations under Article 50 were not postponed in the process.** One transitional rule concerns exclusively **paragraph 2**, that is the machine-readable marking, and only for generative systems already on the market before 2 August 2026. That is a provider duty and concerns practically no deployer. Strictly speaking, paragraph 7 on codes of practice was recast. Beyond the transitional rule mentioned, however, nothing was postponed, and for deployer duties there was never any deferral.

**"In Austria the AI service point act (KISG) is decisive."** No such act exists. The actual legal act is an amendment to the communications authority act and the telecommunications act. It regulates neither market surveillance authorities nor sanctions.

**"In Austria section 3a of the unfair competition act applies."** The Austrian unfair competition act has no section 3a. Breach of law is subsumed there under the general clause. The provision comes from German law.

## 6. What is honestly still unresolved

There is as yet **not a single court decision** on Article 50 in Germany. Nor are we aware of one for Austria. Central terms remain unresolved: when does an image appear deceptively real? Where exactly does the boundary of public interest run? What counts as a sufficient substantive review?

Three conflicting dates circulate for the transitional deadline under paragraph 2: 2 December 2026, 2 February 2027 and 2 November 2026. This point we were able to settle against the Official Journal itself. The Digital Omnibus inserts a new **Article 111(4)** into the AI Act. Under it, providers of generative systems placed on the market before 2 August 2026 must take the necessary measures to comply with Article 50(2) **by 2 December 2026**. The other two dates are earlier negotiating positions since superseded.

What remains open, and we consider it more important, is whether the directly applicable penalties of the regulation can also be enforced where a member state has created neither a competent authority nor a national penalty provision. That concerns Austria and Denmark. We found no reliable answer.

Equally unresolved is whether a **B2B trade blog** falls under public interest. The guidelines' mention of economic and scientific developments speaks in favour; the marketing purpose speaks against. Anyone relying on the all-clear for advertising texts should keep this borderline case in view.

Anyone who gives you a clear-cut answer on these points should be able to explain what they base it on.

## 7. Concrete wordings

Where you conclude that labelling is indicated, these are common formulations:

For texts, following the German competition watchdog's template: **"This text was edited with the help of an AI."** or simply **"AI-generated text"**, set off visually **before** the text

For chatbots: **"You are chatting with an AI assistant."**

For images: a visible note such as **"Created with AI"** in or directly beside the illustration

Two notes on this. The formulation **"AI-supported" is not recommended**, because a human customer-service team can also be AI-assisted at points, which makes the notice meaningless. And the less technically versed your audience, the more you should use terms in the local language rather than English.

Two requirements from Article 50(5) are frequently overlooked, although they apply to all cases under Article 50:

The information must be provided **at the latest at the time of the first interaction or exposure**, meaning at first contact with the content, in a clear and distinguishable manner. A notice found only after use does not satisfy this.

The information must conform to the applicable **accessibility requirements**. This appears in the text of the regulation, not only in the guidelines, and is particularly relevant for public bodies.

The European Commission additionally provides **three official icons**: a base icon, one for fully AI-generated and one for partially AI-modified content. They are available as SVG and PNG and may be used freely **without any attribution requirement**. Their use is voluntary.

An important closing point for this section: **labelling does not cure a misleading practice.** Competition law remains applicable alongside the regulation. Paragraph 6 states expressly that other transparency obligations under Union and national law remain unaffected.

## 8. What follows in practice

The regulation demands less of most companies than the headlines suggest. But it demands something that cannot be made up in a week: **clarity about where AI is involved in your processes at all.**

Anyone who does not know that can neither label nor explain why labelling was unnecessary. And it is precisely that explanation which counts in case of doubt.

Three steps that have proven their worth in practice:

**Take stock.** Where is AI used, by whom, for which content? Without that list everything else is guesswork.

**Define the editorial process.** If you want to rely on the editorial exception, the human must stand at the end and be named.

**Document decisions.** Record why a piece of content was labelled or not. It costs little and pays off later.

## Legal notes

Please read these notes before drawing conclusions from this article.

This article is a **free information offering** and does **not constitute legal advice**. It is not a legal service and does not replace an assessment of your individual case.

The presentation is **general and abstract**. It relates to typical situations and deliberately makes no finding as to which obligations exist in your specific case.

There is **no claim to completeness or accuracy**. The article reflects a section of the legal position.

There is as yet **no case law** on Article 50 of the AI Act. Central terms are unresolved, and interpretation may change.

Reading this article creates **no contract and no engagement**. We assume no duty of examination and do not warrant the accuracy of the information in your specific case.

The article is **not subject to an update service**. It is revised from time to time, without any assurance of currency at any given point.

Liability for **slight negligence is excluded in respect of financial loss**. Liability for intent, gross negligence and death or personal injury remains unaffected. Where obligations are breached whose fulfilment makes proper use of this offering possible in the first place, liability is limited to the foreseeable, typical damage.

This article is an offering of **vellmerk.ai ApS**. It also serves to present our advisory services.

For a reliable assessment of your individual case, please consult a legal adviser of your choice.

**Legal status: 20 August 2026.** The time-sensitive statements on Austria and Denmark were re-checked against the sources on that date.

## Sources and further reading

Regulation (EU) 2024/1689 (AI Act), Article 50, OJ L 2024/1689 of 12 July 2024

Guidelines of the European Commission on the transparency obligations under Article 50, C(2026) 5054 final of 20 July 2026

Code of Practice on Transparency of AI-generated Content, final of 10 June 2026

Regulation (EU) 2026/1744 (Digital Omnibus on AI)

KI-Marktüberwachungs-und-Innovationsförderungs-Gesetz of 22 July 2026 (BGBl. 2026 I No. 223)

RTR, AI service point, transparency obligations and FAQ

Federal Office of Justice (Switzerland), legal baseline analysis for the stocktaking on AI regulation

Act no. 467 of 14 May 2025 on supplementary provisions to the AI Regulation (Denmark)

Wettbewerbszentrale, guidance on labelling AI-generated content, version 2.0 of 29 July 2026

## Frequently asked questions

### Do I have to label a blog post written with AI?

In many cases no, but it depends on the topic. The text obligation in Article 50(4) applies only where three conditions come together: the text is published, it informs the public, and it concerns a matter of public interest. Advertising and product descriptions fall outside the scope as a rule, according to the European Commission's guidelines, as long as they contain no claims relating for example to health, consumer safety or sustainability. Whether a trade blog on economic or scientific topics falls under public interest is, by contrast, unresolved. And even for texts on matters of public interest the obligation falls away where a human has reviewed the substance and a natural or legal person visibly holds editorial responsibility.

### What is the difference between provider and deployer under Article 50?

Providers are the makers of the AI systems, such as OpenAI, Midjourney or Adobe. Deployers are those who use such systems professionally. The regulation distributes the duties very unevenly: paragraphs 1 and 2, including the much-discussed machine-readable marking with watermarks and metadata, address the provider only. For a company using off-the-shelf tools, paragraph 4 is what usually remains, meaning deep fakes and text on matters of public interest. Paragraph 3 is added for those deploying emotion recognition or biometric categorisation. A dual role is possible, for instance where a purchased system is adapted with one's own data and run under one's own name.

### Does the AI labelling duty also apply to companies in Switzerland?

Swiss law contains in principle no labelling duty for AI content; the Federal Office of Justice states this expressly in its legal baseline analysis. It does not follow that Swiss companies are unaffected. The European Commission's guidelines name publishing on the globally accessible internet as a connecting factor for deployers from third countries. A company that itself foresees the dissemination of its content in the EU can therefore fall under Article 50, though only under paragraphs 3 and 4. Distribution channels that are unforeseeable and outside one's own control are not covered.

### What penalties apply for missing AI labelling?

The regulation provides for up to 15 million euros or three per cent of worldwide annual turnover for breaches of the transparency obligations, whichever amount is higher. For small and medium-sized enterprises including start-ups, Article 99(6) reverses this rule: there the lower of the two amounts applies. In practice the situation differs sharply across the four countries considered here. Germany has a competent authority in the Federal Network Agency, Austria has to this day neither designated an authority nor enacted an implementing act, and Denmark currently has no national penalty provision for Article 50. In Austria the practical risk therefore lies not with authorities but with injunction actions under competition law.