AI labelling obligation: what your site must show
The EU AI labelling obligation has applied since 2 Aug 2026. Which AI images, AI texts and chatbots on your website need a label and which do not.
The AI labelling obligation in Article 50 of the EU AI Act has applied since 2 August 2026, but it asks far less of companies with a website than many warnings suggest. You need a visible label for three things: a chatbot, realistic AI depictions of real people, products, places or events that could pass as genuine, and unreviewed AI text on matters of public interest. Generic AI graphics and human-reviewed text need no label. This article reflects the legal position as of October 2026, including the Digital Omnibus, the European Commission guidelines and Germany's implementing act. Below you will find the timeline, the roles under the law, a table by content type, what a label looks like, the real enforcement risk, our take and a checklist.
Key takeaways
- Article 50 of the AI Act has applied since 2 Aug 2026. The Digital Omnibus, Regulation (EU) 2026/1744, did not postpone it. Only providers of older AI tools have until 2 Dec 2026 for machine-readable marking.
- Generic AI illustrations, backgrounds, product descriptions and human-reviewed text need no visible AI label on your website.
- You must label deepfakes: realistic AI images, videos or voices that fake real people, products, places or events.
- A chatbot on your website needs a clear note in the chat window saying an AI is answering. One sentence is enough.
- Fines for transparency breaches: up to EUR 15 million or 3 % of worldwide annual turnover, with the lower figure applying to SMEs (Art. 99(4) and (6)).
- Our decision: no precautionary AI label on everything, targeted labelling where the law requires it.
The short answer
The AI labelling obligation sits in Article 50 of the EU AI Act (Regulation (EU) 2024/1689) and has applied since 2 August 2026. For a company using AI tools on its website, paragraph 4 matters most: deepfakes must be disclosed, meaning AI images, video or audio that resemble real people, objects, places or events and falsely appear authentic. The same applies to AI text published to inform the public on matters of public interest, unless a human has reviewed it and someone holds editorial responsibility.
Machine-readable marking of AI content, for example through watermarks or metadata, is the duty of the tool providers, not yours. A chatbot on your website must make clear that an AI is answering.
For most marketing images on company websites, such as illustrations, backgrounds and key visuals without real people, and for text someone on your team has reviewed and approved, there is no visible labelling duty. Labelling everything anyway does not fulfil any additional obligation.
When did the AI labelling obligation start?
Article 50 has applied since 2 August 2026. In July 2026 the EU adopted the Digital Omnibus, an amendment package for the AI Act that moved several deadlines. The transparency duties were not among them. The main postponement concerns high-risk AI, for example in recruitment or credit scoring.
When did the AI labelling obligation start?
| Date | What happens | Relevance for your website |
|---|---|---|
| 2 Feb 2025 | Bans on certain AI practices apply | Low, covers manipulation and social scoring |
| 20 Jul 2026 | European Commission publishes Article 50 guidelines and a code of practice | Interpretation aid, not legally binding |
| 24 Jul 2026 | Digital Omnibus (Reg. (EU) 2026/1744) published, in force since 27 Jul 2026 | Changes nothing about your Article 50 duties |
| 29 Jul 2026 | German AI market surveillance and innovation act (KI-MIG) in force, Federal Network Agency as lead authority | Contact point and complaints body in Germany |
| 2 Aug 2026 | Article 50 applies in full, including to deployers | Chatbot notice and deepfake labelling apply from here |
| 24 Sep 2026 | European Commission publishes optional EU icons for labelling | Ready-made symbols if you need to label |
| 2 Dec 2026 | Transition period ends for machine-readable marking by AI tools on the market before 2 Aug 2026 | Applies to tool providers, not a deadline for you |
| 2 Dec 2027 | Duties for stand-alone high-risk AI (previously 2 Aug 2026) | Only for high-risk use |
The 2 December 2026 date is often misread. According to recital 38 of the Digital Omnibus, the four-month transition applies only to providers of generative AI systems that placed them on the market before 2 August 2026. For you as a website operator, it is no reprieve.
Provider or deployer: who carries which duty?
The AI Act distinguishes two roles. Providers develop AI systems and place them on the market under their own name, such as OpenAI, Google or image AI vendors. Deployers use AI systems in a professional capacity (Art. 3(4)). That is your company as soon as your marketing team creates images with Midjourney or writes copy with ChatGPT. According to the European Commission, private, non-professional use is outside the regulation.
- Para. 1, providers: design AI systems that interact directly with people so users know they are talking to an AI.
- Para. 2, providers: mark AI output in image, audio, video and text in a machine-readable way, for example via C2PA, an open standard for provenance data in a file's metadata.
- Para. 3, deployers: inform people when emotion recognition or biometric categorisation is running. Rare on normal websites.
- Para. 4, deployers: disclose deepfakes and label AI text on matters of public interest unless it has been human-reviewed and editorially owned.
With chatbots the role is not always clear-cut. Formally, the paragraph 1 duty falls on the provider. If you run a bot under your own name on your website, you may count as a provider yourself, depending on the setup. You do not need to settle that question: one sentence in the chat window meets the duty, whoever carries it.
Do I have to label AI images on my website?
Only if the image is a deepfake. Article 3(60) defines it like this: the content resembles existing persons, objects, places, entities or events and would falsely appear authentic to the viewer. The Commission breaks this into three tests that must all be met: close resemblance, a real-world original and a false appearance of authenticity. A deepfake is not a verdict on image quality. It is a question of deception.
The Commission guidelines of July 2026 get specific with examples. A real product in front of an AI-generated background is not a deepfake as long as the ad does not mislead. Content that defies the laws of physics or is clearly fantastical is also out of scope. An AI product image that presents the product's appearance, features or use differently from reality, on the other hand, counts as a deepfake. So does an AI-generated depiction of a celebrity in advertising.
For websites, that means the hero visual with an abstract 3D scene, the illustration in a blog article and the background behind your real product photo need no label. An AI image of a client project, an event or a client that never happened does. AI portraits of your actual team are an open question: the person exists and really works for you. Whether such a portrait "falsely appears authentic" in the sense of the regulation has not been settled by the Commission or by any court. Expert commentary also disagrees on photorealistic AI people with no real-world model. For testimonials, label them or skip them entirely, because German unfair competition law, the UWG, also applies there: an invented client is misleading with or without the AI Act.
What about AI text on the website?
For text, the duty applies only in a narrow case: when the text is published to inform the public on matters of public interest. The Commission lists politics, public administration, justice, public health, environmental protection and consumer safety as examples. According to the guidelines, advertising copy and product descriptions are not covered as long as they make no claims about health, consumer safety or sustainability.
Even when a text falls under the rule, the label is not required if two conditions are both met. The text has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for publishing it. The guidelines require a genuine substantive review, at minimum a fact check. A quick glance before hitting publish does not count.
For your service pages and expert articles, that means: if someone on your team reviews the content, checks the facts and approves it, you need no label. That is the standard we recommend for SEO copy anyway. The risky cases are auto-generated pages without sign-off, for example mass-produced guides on health or legal topics.
Which content on your website needs a label?
The table sums up the typical cases. It is based on Article 50 of the AI Act and the Commission guidelines of July 2026. Where interpretation is still open, the table says so.
Which content on your website needs a label?
| Content on the website | Visible label needed? | Why |
|---|---|---|
| Generic AI illustration, abstract graphic, background | No | Shows nothing real, does not look like a genuine photo |
| Real product photo in front of an AI-generated background | Usually no | Not a deepfake per the guidelines, as long as nothing about the product is misrepresented |
| AI product image showing your product differently from reality | Yes | Guidelines treat it as a deepfake, plus misleading advertising under the UWG |
| AI image, video or voice of clients, events or reference projects that never existed | Yes | Deepfake under Art. 50(4), invented references are also unfair competition |
| AI portrait of real staff with their consent | Grey area | Person and role are real, whether the portrait "falsely appears authentic" is unsettled |
| Photorealistic AI person with no real-world model | Grey area | Interpretation disputed, label in testimonial contexts |
| Service page, product description, ad copy written with AI | No | Not text on matters of public interest |
| AI-assisted expert article, human-reviewed and approved | No | The Art. 50(4) exemption applies |
| Unreviewed AI text on health, law or politics | Yes | Art. 50(4), no editorial responsibility |
| Chatbot or AI agent in the chat window | Yes | Art. 50(1), notice at first contact at the latest |
| Translation or spell-checking with AI | No | Assistive editing, no duty according to IHK Cologne |
What does a label look like when you need one?
Article 50(5) requires the notice to be clear and distinguishable and to appear at the latest at first contact with the content. It must also be accessible. That links the AI duty to a second obligation for websites, Germany's accessibility act, the BFSG. According to the Commission, notices in your terms or purely machine-readable marks are not enough. A line in your legal notice does not reach visitors at first contact either.
- Image or video: label directly in the content, at the start for videos. The Commission's EU icons show the letters "AI" in three variants: basic, fully AI-generated, partly AI-modified. Using them is optional.
- Text: label above the text or near the headline, for example "This text was created with AI and has not been editorially reviewed."
- Audio: a short spoken notice at the start.
- Chatbot: one sentence in the chat window before the first reply, for example "You are chatting with an AI assistant. For personal advice, reach our team via the contact form."
One detail on the icons is worth knowing: according to the Commission, the basic icon performed better in testing when a short text label sat next to it. Build the label into your website's design system instead of overlaying it on images later with a plugin. If you are planning a relaunch anyway, add it to the component list.
How realistic are fines and cease-and-desist letters?
For breaches of Article 50, Article 99(4) of the AI Act provides for fines of up to EUR 15 million or 3 % of worldwide annual turnover, whichever is higher. For small and medium-sized companies, Article 99(6) applies the lower of the two. The widely quoted EUR 35 million or 7 % only applies to prohibited AI practices under Article 5, not to missing labels.
In Germany, the KI-MIG, the national act implementing the AI Act, has been in force since 29 July 2026. The Federal Network Agency is the central supervisory and complaints body and runs an AI service desk for smaller companies. Up to our editorial deadline, we found no publicly documented fines for missing AI labels. The agency describes itself as an enabler and point of contact.
The more realistic risk is a cease-and-desist letter. Competitors and associations can act against misleading advertising under the UWG, and several law firms consider Article 50 a market conduct rule whose breach can be pursued that way. That risk concentrates on the cases where the duty actually exists: embellished AI product images, invented testimonials, AI photos of supposed clients. For a generic illustration, there is nothing to pursue.
Where the duty does not apply: the gaps in the law
Article 50 sounds sweeping but is narrowly drawn. For most content on a company website it either does not apply or does not require anything of you. As things stand, these cases fall outside it:
- Generic AI visuals: illustrations, backgrounds, abstract 3D scenes and icons show nothing real, so they are not deepfakes.
- Advertising and product copy: the text duty only covers text on matters of public interest. Service pages, product descriptions and campaign copy are not included according to the guidelines.
- Reviewed text: even an expert article on a public topic needs no label if a person has reviewed it and your company holds editorial responsibility.
- Watermarks and metadata: machine-readable marking under paragraph 2 is the tool providers' duty. You do not have to retrofit metadata.
- Obviously artistic content: for clearly artistic, satirical or fictional content, a disclosure that does not hamper the display is enough.
- Assistive editing: according to IHK Köln, translation and spell-checking with AI do not trigger a label.
- Existing images: commentators disagree on whether older AI images must be labelled retroactively. There is no clear answer yet.
What remains is narrow: chatbots, deceptively real depictions of things that never existed, and unreviewed AI text on public topics.
Our take
We discussed the legal position with lawyers specialising in IT law and compliance. Their assessment as things stand: for typical website content, the labelling duty is hardly enforceable in practice, because authorities cannot tell whether an illustration or a reviewed text was made with AI. And anyone who marks every image and every text as "AI-generated" as a precaution gives away marketing impact without gaining anything legally.
That matches the text of the law. For generic visuals and edited text, there is no visible labelling duty. An AI label under every hero image and every paragraph is overblown panic, not a compliance gain. On a B2B website it signals mass production rather than care, and that hurts conversion.
Those are the official rules. In practice, hardly anyone follows them: on the company websites we see in projects and competitor analyses, we almost never find an AI label, not even where AI images are clearly in use. That is no free pass, but it shows how far the public alarm and everyday reality are apart.
Where the duty clearly applies, we label. Chatbots get a notice, which costs one sentence. We do not use AI images of clients, projects or events that never existed, neither on a corporate website nor in campaigns. We do not publish unreviewed AI text, and not only because of Article 50.
Not legal advice: this is our assessment as of October 2026. For specific borderline cases, such as photorealistic AI people in campaigns, a short call with a specialist law firm is worth it.
Do you need to change anything on your website?
For most company websites the effort is small. What matters is which AI content you actually use.
Yes, if
You need a label or a change as soon as one of these points applies to your website.
- Your website runs a chatbot or AI assistant without a clear notice in the chat window.
- You show AI-generated images of clients, locations, events or reference projects that do not exist in that form.
- Your AI product images show products nicer, bigger or different from what they are.
- You publish AI text on health, legal or consumer topics without expert review.
No, if
You do not need to change anything if your AI use falls into these categories.
- You use AI for illustrations, abstract visuals or backgrounds behind real product photos.
- Your copy is written with AI support but reviewed and approved by your team.
- Your website has no chatbot and no realistic AI depictions of real people.
- You use AI only for translation, proofreading or photo retouching such as exposure and cropping.
Checklist for website owners
For a typical B2B website these steps take a few hours. Start with the inventory, and the rest follows from it.
Take stock of AI content:
List where your website uses AI images, AI videos, AI text and AI tools such as chatbots.
Apply the deepfake test:
Check every AI image and video for the three features: real-world resemblance, an existing model and a false impression of authenticity.
Add a chatbot notice:
Add one sentence to the chat window that is visible before the first reply.
Define a sign-off process for text:
Decide who reviews AI text for substance, checks facts and takes responsibility for publishing. Put it briefly in writing.
Review critical images:
Swap AI images of supposed clients, events and projects for real shots, or add a visible label with EU icon and text. AI portraits of your real team are a grey area, make a deliberate call there.
Make the label accessible:
Build the notice as a component with alt text so screen readers can read it.
AI itself remains a tool we use in projects every day, from building a website with AI to optimising for AI search systems. How to become visible there is covered in our articles on generative engine optimization and llms.txt. If you want to anchor labelling properly in your design right away, it belongs in a web design project or a website relaunch.
Frequently asked questions
The AI labelling obligation is set out in Article 50 of the EU AI Act and has applied since 2 August 2026. Companies must disclose deepfakes and unreviewed AI text on matters of public interest. Chatbots must make clear that an AI is answering. AI tool providers must also mark their output in a machine-readable way.
Only if they are deepfakes, meaning they realistically show something that exists and falsely appear authentic. An abstract AI illustration or an AI background behind a real product photo needs no label. An AI image of a client or event that never existed, or an AI product image that makes your product look better than it is, does.
Yes. Under Article 50(1), users must be able to tell they are interacting with an AI unless that is obvious. The Commission reads this exception narrowly. One sentence in the chat window, visible before the first reply, meets the duty. A note buried in your terms of service does not.
No. Article 50(5) requires a clear and distinguishable notice at the latest at first contact with the content. According to the Commission, notices hidden in terms or purely machine-readable marks are not sufficient. The notice belongs directly on the image, video or text, or in the chat window.
On 2 December 2026 a transition period from the Digital Omnibus ends. It applies only to providers of generative AI tools that placed their system on the market before 2 August 2026, and only to machine-readable marking. For website operators it is no reprieve: your duties have applied since August.
The Federal Network Agency is the central authority under the KI-MIG, in force since 29 July 2026. It takes complaints and runs an AI service desk. Fines for transparency breaches reach EUR 15 million or 3 % of annual turnover, the lower figure for SMEs. Cease-and-desist letters from competitors under the UWG are a second route.
Next step
Want to know which AI content on your website needs a label and which does not, without marking everything as a precaution?
