Labeling AI content on your website: Art. 50 of the AI Act in practice
Labeling AI content under Art. 50 AI Act: which images, videos and texts on your website to label, what you can skip, ready-made captions and a checklist.
More and more businesses create graphics, product photos and blog posts with the help of AI. Since 2 August 2026, some of this content must be clearly labeled, because Art. 50 of the AI Act has started to apply. It is easy to fall into one of two extremes: labeling nothing, or adding “AI” under every color correction. Both cause trouble for a business.
In this guide, you will find a table showing what needs to be labeled on a business website and what doesn’t, ready-made captions to copy, a code example and a checklist. I base it on the text of the regulation and on the European Commission’s guidelines of 20 July 2026. I am not a lawyer, so this is not legal advice. As of 30 October 2026.
Art. 50 of the AI Act in a nutshell: who, what and from when
The AI Act is Regulation (EU) 2024/1689. Its Art. 50 concerns transparency and has applied since 2 August 2026. The EU Digital Omnibus of July 2026 did not move this date. It only gave providers of generators that were on the market before 2 August until 2 December 2026 to mark content technically.
The provision divides obligations between two parties:
- Providers of AI tools, that is, the makers of image, text and voice generators, must mark generated content in a machine-readable way, for example in metadata. This isn’t your job.
- Deployers of AI, that is, every business that uses AI professionally, must disclose two things: deepfakes and AI-generated text on matters of public interest, if nobody has checked its substance.
Private use is excluded, but a graphic for a business website is already professional use. According to the guidelines, if you commission an ad from an agency and don’t decide whether and how it uses AI, the agency is the deployer. However, when you generate graphics yourself, or an employee or a freelancer does it on your instructions, the obligation is yours.
A separate obligation concerns chatbots: the customer must know right away that they are talking to AI. I describe this in the guide AI chatbot and the AI Act. The assistant on my website says so from the first message. If you are only planning an AI chatbot for your website, I will build this notice in from the start.
In Poland, the supervisory authority is the Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji (Commission for the Development and Security of Artificial Intelligence), established by the ustawa o systemach sztucznej inteligencji of 3 July 2026 (Poland’s Act on Artificial Intelligence Systems). The provisions of this act on inspections and penalties have applied since 28 October 2026, and you can submit a complaint to the Commission. For a breach of Art. 50, the AI Act provides for fines of up to EUR 15 million or 3% of annual worldwide turnover, whichever is higher. For small and medium-sized businesses, the upper limit is the lower of these amounts.
What needs to be labeled and what doesn’t
The table covers the most common content on a small business website. The assessment always depends on the specific case, but the examples come from the Commission’s guidelines.
| Content on the website | AI Act label | Why |
|---|---|---|
| Color correction, cropping, removing dust or the background in a product photo | No | Ordinary editing that doesn’t change the meaning of the photo |
| A real product on an AI-generated background | Usually not | If the background doesn’t mislead about how the product looks and works |
| A product that looks better than in reality thanks to AI | Yes | It is a deepfake, and on top of that there is a risk of misleading the customer |
| A realistic “photo” of a crew at work or of a project that never happened | Yes | It shows a service that could have happened but didn’t |
| An AI avatar or the owner’s cloned voice in a video | Yes | A realistic person whose statement isn’t real |
| A cartoon, icons, an illustration in an unrealistic style | No | Nobody will take it for real |
| A product description, offer or ad text written with AI | Usually not | Advertising is out of scope, as long as it doesn’t contain claims, e.g. about health, safety or the environment |
| An article about health, new legislation or consumer safety, published without a substantive check | Yes | Text on matters of public interest |
| The same article checked by a person who knows the subject and is responsible for it | No | Exception for editorial control |
Photos and graphics: when an edit becomes a deepfake
A deepfake in the AI Act isn’t just a fake recording of a politician. It is image, audio or video content generated or manipulated by AI that resembles existing or realistically possible persons, objects, places or events and could pass as authentic. The Commission also counts the depiction of a service being performed as an event. That is why a realistic image of “our crew during installation” generated from scratch needs a caption.
The guidelines list minor changes that usually don’t turn an image into a deepfake:
- cropping, color correction, brightening and sharpening,
- removing dust, red eyes or random passers-by in the background,
- changing the background purely for aesthetics, combining real products and scaling photos in an ad.
The guidelines consider significant changes to include, among others, adding or removing objects in a way that changes the meaning of the photo, changing a person’s figure or face, and photomontage that alters the image of people, things or events. In that case, assess whether the image could pass as real.
Context and audience also count. According to the guidelines, mice arguing in human voices about the best cheese in a cheese producer’s ad aren’t a deepfake, but a realistic AI influencer testing a real product is. If your website is viewed by older people or children, they are easier to mislead, so make your assessment with them in mind.
Remember that a label doesn’t fix dishonest content. If AI shows a product as nicer than it is, a caption won’t protect you from the charge of misleading the customer. On your website, show real photos of your projects, and use AI images only as visualizations. I give a template note for visualizations in a proposal in the guide on AI bathroom visualization.
AI texts: blog, descriptions and news
Product descriptions, offers and ad texts written with the help of AI don’t require an AI Act label. The guidelines explicitly give them as an example of text that is out of scope. However, they make a reservation for claims about, for example, health, safety or the environment, so be careful with such texts.
Articles that inform about matters important to the public are different: health, legislation, consumer safety, the environment or the economy. A medical practice’s post about a new recommendation for patients, or an accounting firm’s post about tax changes, generated and published without checking, most likely needs to be labeled.
You don’t need a label when you meet two conditions at once:
- The text was checked by a person who knows the subject. This isn’t about fixing typos. The guidelines require substantive verification, including fact-checking.
- Someone is responsible for the publication. That person’s name or role and contact details should be easy to find, for example under the article or in the website’s legal information.
If, after approval, you run the text through AI again and its content changes, the exception no longer applies. The last word must belong to a human. In a small business, a simple rule is enough: every post has an author, a reviewer and an update date.
How to label: placement, icons and ready-made captions
The information must be clear, distinguishable from the rest of the content and appear at the latest when someone first sees the material (Art. 50(5)). According to the guidelines, a mention in the terms and conditions or in instructions isn’t enough. Put the caption next to the image, and at the beginning of a video. If viewers can start watching from the middle, repeat it or keep a permanent badge in the corner.
Hidden markings added by the generator (metadata or an invisible watermark) aren’t enough, because people can’t see them. Don’t remove them, though. Google Merchant Center requires AI images to keep the IPTC TrainedAlgorithmicMedia tag, and titles and descriptions written by AI must be submitted in the structured_title and structured_description attributes. Check that your image compression plugin doesn’t strip the metadata.
The Commission has prepared three free EU icons for labeling AI-generated content: a basic one, “fully AI-generated” and “partially AI-modified”, as SVG and PNG files. Using them is voluntary and doesn’t by itself settle compliance with the law, so add a short text next to them.
Ready-made captions to copy:
Illustration generated by AI.
Photo modified with AI: living room background added.
Visualization generated by AI. The final result may differ.
The character and voice in this video were generated by AI.
Article prepared with the help of AI.
On a website, the simplest way is to add a caption under the image. A screen reader will read it together with the graphic:
<figure>
<img src="salon-sofa.webp" alt="Gray sofa in a bright living room, visualization generated by AI">
<figcaption>Illustration generated by AI.</figcaption>
</figure>
Don’t hide the information in the title attribute, which only appears when you hover the mouse over the image. Nobody will see it on a phone. On social media, use the platform’s built-in AI label: for very large platforms such as Facebook, Instagram or TikTok, the guidelines allow you to rely on it if it is clearly visible.
Checklist: sorted in 30 minutes
- List the places where your business uses AI: website, blog, store, PDF offers, social media, chatbot.
- Go through the graphics and videos added since 2 August 2026 and caption those that match the “Yes” rows in the table.
- Check that your chatbot introduces itself as AI in its first message.
- Decide who checks the substance of blog posts, and name that person under the articles. Label posts without such verification.
- Add to your contract with a graphic designer, an agency or a freelancer that the contractor states which materials were created with AI and delivers them labeled.
- Write the rules down in a short guide for your team: what we label, with what caption and where.
- Older content: you don’t have to retroactively label deepfakes generated before 2 August 2026, although the Commission encourages it. Text generated earlier but published after that date is already subject to the new rules.
- Once a quarter, look at the Commission’s guidelines, because the Commission has announced that it will update them.
Summary
On a business website, you label realistic AI images, videos and voices that someone could take for real, as well as unchecked texts on public matters. You don’t have to label ordinary photo editing, product descriptions or cartoon-style graphics, but a short caption “Illustration generated by AI” costs nothing and builds trust.
If you want this off your plate, as part of the AI Act package from PLN 490 net I will label AI graphics, videos and texts where the law requires it, add an AI notice to the chat and prepare instructions for your team. I will prepare new graphics and texts, with the right captions from the start, as part of the texts and graphics service. Write to me: I reply within 24 hours, and new clients get a 20% discount on their first order.
Frequently asked questions
Usually not. In its guidelines on Art. 50, the European Commission treats color correction, cropping, changing the background for aesthetic reasons or upscaling a photo as minor changes that don’t turn an image into a deepfake. A label is needed when AI changes how the product looks, but then misleading the customer is also a problem.
Only if it concerns matters of public interest, for example health, legislation or consumer safety, and nobody has checked its substance. If the text was verified by a person who knows the subject and the business takes responsibility for it, the AI Act label isn’t required. An ordinary guide, for example on caring for a terrace, usually doesn’t concern matters of public interest.
No. Hidden markings are the tool provider’s obligation, and people can’t see them. A business that publishes a deepfake must add a visible caption or icon next to it. Don’t remove the metadata, though, because Google Merchant Center, for example, requires it.
There is no obligation to retroactively label deepfakes generated before 2 August 2026, although the European Commission encourages it. Text on public matters is different: if AI wrote it earlier but you publish it after that date, the new rules apply.
In Poland, the supervisory authority is the Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji (Commission for the Development and Security of Artificial Intelligence), to which you can submit a complaint. For a breach of Art. 50, the AI Act provides for fines of up to EUR 15 million or 3% of annual worldwide turnover, and for small and medium-sized businesses the lower of these amounts applies. The authority takes into account, among other things, the seriousness of the breach and the business’s cooperation.
Related services
Have a similar project?
Describe what you need and I’ll send a quote within 24 hours. New clients get 20% off their first project.
