AI labelling requirements: what applies now?

Since August, the same worry has been doing the rounds in many teams: does every text that ChatGPT helped to write now need an AI stamp? Every newsletter, every LinkedIn post, every product description? Some guides suggest exactly that. The answer is more relaxed, but it is not a free-for-all: the labelling obligation in the EU AI Act targets very specific cases, and they can be clearly identified.

Article 50 of the AI Act has applied since 2 August 2026. What it requires of small businesses is now fairly clear, because the European Commission published detailed guidelines on it in July 2026. Here is the overview, as of October 2026.

Key points at a glance

• Since 2 August 2026, the transparency obligations under Article 50 of the EU AI Act have been in effect. In Germany, supervision is the responsibility of the Federal Network Agency (Bundesnetzagentur).
• Advertising copy, product descriptions and ordinary company posts do not need to be labelled, even if AI helped to write them. Exception: if they contain statements on topics such as health, safety or sustainability.
• Texts on matters of public interest must be labelled, unless a human has reviewed their content and someone visibly holds editorial responsibility.
• Realistic AI images, videos and voices that could pass as real are deep fakes and must be recognisable as AI. This also applies to invented but realistic people.
• Chatbots must say at the start that they are an AI. Who is responsible for this depends on whether you build the bot yourself or integrate a ready-made one.
• Fines go up to €15 million or 3% of annual turnover; for small and medium-sized enterprises, whichever amount is lower applies.

What has been in effect since 2 August 2026

Article 50 of the AI Act contains four transparency obligations. Two of them apply to the companies that make AI tools, two to the businesses that use AI. The AI Act calls them providers and deployers.

For most small businesses, paragraph 4 is the part that matters. If you use ChatGPT, Midjourney or a voice tool for work, you are a deployer. Purely private use is exempt; anything employees create on behalf of the business falls under the business’s obligations.

Are you a provider or a deployer?

Usually a deployer. You use ready-made AI tools for your work; the company that makes the tool is the provider and takes care of the technical marking. You only become a provider if you develop an AI system yourself, or have one developed, and use it under your own name, for example your own chatbot. You can hold both roles at the same time.

One detail takes the pressure off many businesses: according to the Commission, if you hire an agency to run a campaign and do not decide yourself whether and how AI is used in it, you are not a deployer. The labelling obligation then lies with the agency. It is still wise to put this in writing in the agency contract.

AI texts: when you have to label them

An AI text only has to be labelled if three things come together. It is published, meaning made available to an open group of readers. It is meant to inform, meaning to convey knowledge, facts or opinions. And it concerns a matter of public interest, such as politics, law, health, the environment, consumer safety or economic developments that are being debated in public.

The Commission gives examples on both sides. Not covered are novels, advice letters to individual customers, and AI-edited advertising and product descriptions, as long as they do not contain statements on topics such as health, consumer safety or sustainability. Covered, on the other hand, are things like a guide article on diet for people with a particular illness or the annual report of a listed company. Internal documents and emails to individuals are never “published”.

For small businesses, this means: if you review AI drafts for accuracy before publishing them and name a responsible person, you will generally not have to label your specialist blog or newsletter. The review should be traceable, for example with the approval date and a name.

AI images, videos and voices: when it is a deep fake

Under the AI Act, a deep fake is AI-generated or manipulated image, audio or video content that resembles existing persons, places, objects or events and would falsely appear to be authentic. Two points surprise many people:

• Invented people count too. A photorealistic AI person in the style of a stock photo can be a deep fake, even if there is no real person behind it. The same goes for a voice clone or a realistic avatar speaking on behalf of your management.
• No intention to deceive is required. What matters is whether the audience could take the content to be real.

Minor edits are not deep fakes: colour correction, cutting a product out of its background, or swapping a background for purely aesthetic reasons. Illustrations, comics or obviously surreal images are not covered either. It gets tricky when an image makes your product look better than it really is. According to the guidelines, that is a deep fake and may also count as misleading advertising. For evidently artistic or satirical works, a discreet notice that does not hamper the enjoyment of the work is sufficient. The Commission interprets this exception narrowly; ordinary advertising usually does not fall under it.

Chatbots on your website

If you use a chatbot to talk to customers, it must identify itself as AI at the start. The “obvious” exception rarely applies: to many users, a bot in a support window looks like a human on the other end. A good standard is a first chat message such as “Hello, I’m an AI assistant”. According to the Commission, vague phrases such as “This website uses AI” or simply calling the bot “Assistant” are not enough.

Formally, the obligation lies with the provider of the chatbot. If you integrate a ready-made product without changes, that is the company that makes it. If you build the bot yourself or have it developed for you, it is you. What is still unclear is the case where you run a standard widget under your own name. Either way, the safe solution is the same: make sure the notice appears, no matter who is legally responsible for it.

Do I have to label this? A decision guide

How to label correctly

The label must be perceptible without technical aids. The invisible marking that an AI tool writes into the file does not replace your notice. Three rules from the Commission’s guidelines:

1. Visible and unambiguous, for example “AI-generated image” directly on the image, or “This video contains an AI-generated voice” at the start.
2. At the latest at the time of first exposure, so not only at the end of a video and not hidden away in your terms and conditions or terms of use.
3. For each piece of content, not once as a blanket statement for the whole website.

On Instagram, Facebook, TikTok or YouTube, you can use the platform’s own labelling feature, as long as the notice is clear and visible directly on the content. Regardless of the AI Act, the platforms already require realistic AI content to be labelled. If you leave the label off there, you risk reduced reach or removal, even though that is not a fine.

Who enforces the rules and what you risk

Since 29 July 2026, the German AI Market Surveillance Act (KI-MIG) has governed supervision in Germany. As a rule, the Bundesnetzagentur is responsible; it also runs a central complaints office. Breaches of Article 50 can be punished with fines of up to €15 million or 3% of worldwide annual turnover. For small and medium-sized enterprises, whichever amount is lower applies. A business with €2 million in turnover would therefore face a maximum of €60,000. No reason to panic, but not a trivial sum either.

Then there is competition law: the Wettbewerbszentrale (Germany’s Centre for Protection against Unfair Competition) takes the view that breaches of the AI Act can also be pursued with a formal warning letter (Abmahnung). There is no case law on this yet.

The most common mistakes

• Labelling everything across the board. A “Made with AI” stamp under every text is not legally required and dilutes the notices that actually matter.
• Relying only on the tool’s metadata. Deep fakes need a visible notice; the technical marking alone is not enough.
• Treating approval as a formality. The editorial exception only applies if there is a genuine expert review that includes a fact check. A quick glance from the intern is not enough.
• Underestimating realistic AI people. Even an invented, photorealistic person in an ad can be a deep fake.
• Forgetting the chatbot. The website widget has been running for months, and the AI notice is missing.
• Panicking and labelling old content after the fact. You do not have to mark deep fakes created before 2 August 2026 retrospectively. For texts, the publication date is what counts: anything that only goes online after that date is subject to the new rules.

How Klaus helps you with labelling

The question “Do I have to label this?” does not come up just once. It comes up again with every tool and every use case. Klaus turns it into a fixed rule: in the free onboarding chat, Klaus records your AI tools along with what you use them for, for example ChatGPT for blog drafts, an image generator for social media, a chatbot in customer service. From this, Klaus creates your AI policy with a dedicated chapter on transparency, labelling and copyright. It sets out clearly for your team which content is reviewed before publication and which content is labelled.

If you use AI in dealings with customers, for example as a chatbot or for image and video content, Klaus systematically checks in the assessment project which transparency obligations under Article 50 apply and creates a task for implementing them. In the team training, your employees work through the cases from this article using your own tools.

Conclusion

The AI labelling obligation is narrower than its reputation suggests. You can write advertising copy, product descriptions and ordinary company posts with AI without stamping them. What you do have to label are realistic AI images, videos and voices, your chatbot, and texts on matters of public interest that nobody with the relevant expertise has reviewed. If you review AI drafts carefully, name a responsible person and add a clear notice to realistic media, you are on the safe side. You can organise all of this in a morning, and it belongs in your AI policy as a rule.