Are satire and caricature exempt from labelling?

No, and this is the most common mistake. For evidently artistic, creative, satirical or fictional work the duty is reduced, not removed. You still have to disclose that the content is generated, but you may do it in a way that does not hamper the display or enjoyment of the work, for example in the caption or the credits.

Does this apply to a business based in the United States?

It can. The AI Act reaches providers and deployers outside the EU where the output of the system is used in the Union. If your site addresses customers in the EU, plan for it. If you only ever sell inside the US, this specific rule does not bind you, though disclosing anyway is cheap and builds trust.

Where exactly does the notice have to sit?

Article 50(5) requires the information to be clear and distinguishable, at the latest at the time of the first exposure, and to meet accessibility requirements. For an image on a web page that means next to the image, not buried in a policy page and not behind a click. A badge on the image plus a line underneath does the job.

What does a breach cost?

Breaches of the transparency duties carry fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. In Germany the Federal Network Agency runs market surveillance under the national implementing act.

Labelling AI images on a website: the EU rules since August 2026

Q: Do I have to label every AI image on my website?

No. The duty in Article 50(4) of the EU AI Act hangs on the word deepfake. That means image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic. An image that obviously reads as a drawing does not qualify. A photorealistic office scene does.

Article 50 of the EU AI Act, the part about transparency for artificially generated content, has applied since 2 August 2026. If you put AI images on your website, you now need to know when a disclosure belongs next to them.

Short version: not every AI image needs a notice, but far more of them do than most people assume. And the exemption everyone quotes for satire is not an exemption from labelling at all. It only softens how the label may look. That single point is wrong in a lot of guides.

Why a US business should care

The AI Act is European law, so the obvious question is whether it reaches you at all. It can. The regulation covers providers and deployers established outside the Union where the output produced by the system is used inside the Union. In plain terms: if your site is aimed at customers in the EU, plan for these rules. If you sell only inside the US, this particular obligation does not bind you, but the practice below is still worth adopting. Disclosure costs one line and buys trust.

Two roles, two duties

The regulation separates providers from deployers, and that split decides which part is your problem.

A provider develops an AI system and puts it on the market. OpenAI, Google, Midjourney. Under Article 50(2) they have to mark their outputs in a machine-readable format, through metadata such as C2PA or a watermark. That happens inside the tool, so it is not your job.

A deployer uses an AI system under its own authority, outside a purely personal activity. Generate an image for your company website and you are a deployer. Article 50(4) is the paragraph that applies to you, and it turns on one word.

The word everything hangs on: deepfake

Article 3(60) defines a deepfake as

AI-generated or 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.

Two conditions, and both have to be met: it resembles something that exists, and it would pass as real. That gives you a workable rule of thumb:

Photorealistic means disclose. A generated team in a meeting room, a generated product shot, a generated storefront. The fact that the people do not exist is no help: the wording covers objects, places and events too.
Obviously drawn usually does not. A flat icon scene, a process diagram, an illustration with visible outlines. Nobody mistakes that for a photograph.
In between there is a narrow strip. The closer an illustration gets to a photograph, the more likely the duty applies. The doubt rule belongs to that strip only. For anything clearly on one side of the line, there is no doubt to resolve.

Three examples

A photorealistic AI image with no disclosure at all, marked in red as the wrong example

Not like this. The image is meant to read as a photo of your own team and carries no notice. The caption even asserts a real situation. This is exactly the case Article 50(4) was written for.

The same image with a badge on the picture and a note in the caption, marked in green as the right example

Like this. A badge on the image and a line underneath. Together that is more than the law demands: the line on its own is enough. What matters is that the notice sits with the image, not in a legal page in the footer. If you want it quieter, there are four ways further down.

An illustration that obviously reads as a diagram, no disclosure, marked in green

No duty. This one is not a deepfake. It resembles nothing that exists and nobody would take it for a photograph, so the first of the two conditions in Article 3(60) is already missing. You do not have to label this, and you should not: a notice where none is needed devalues the notices where one is.

Which images need a disclosure

Image type	Example	Disclosure	Why
Photorealistic AI image of people, places or products	A generated team in a meeting room	Yes	Resembles existing persons, objects and places and would pass as real. Exactly the case in Article 3(60).
AI image showing a specific real person	A generated portrait of a client	Yes	The same case, sharpened. Personality and likeness rights apply on top, independent of the AI Act.
Real photo, altered in substance with AI	A person removed, a background replaced	Yes	The picture now asserts something other than what was in front of the camera, and still reads as a photograph.
Real photo, only technically processed	Brightened, cropped, denoised, upscaled	No	The picture is still what it was. Standard editing does not change the claim.
Obviously an illustration or drawing	An icon scene with visible outlines	No	Nobody takes it for a photograph, so the second condition is missing.
Diagram, icon, chart	A four step process	No	Same, even when an AI drew it.
Logo, pattern, abstract graphic	A generated background gradient	No	Resembles nothing that exists.
Photorealistic but plainly impossible	An elephant on a rooftop, rendered photographically	Borderline	Photorealistic yes, believable no. In genuine doubt, disclose.
AI image inside an evidently satirical or artistic work	A caricature	Yes, softened	The duty stays, only the form gets gentler. See the next section.
Stock photo or your own photo, no AI	A photo from a photographer	No	No AI involved.

Satire is not exempt

The line you read everywhere is that caricature and satire need no label. That is not what the text says.

Article 50(4), first subparagraph, says that where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the transparency obligation is limited to disclosing the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work.

So the duty stays, only the form gets gentler. Instead of a badge across the picture, the notice may live in the caption, in the credits or in accompanying information. On top of that, the fit into one of those categories has to be evident case by case. Purely informative or promotional content does not qualify, and where the two mix, the informative character wins.

For a company website the practical answer is simple: do not build on the satire exemption. Your site informs and promotes.

AI text is the smaller problem

For text, Article 50(4), second subparagraph applies, and it is narrow. It covers text published with the purpose of informing the public on matters of public interest. The duty falls away where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.

Translated: a services page, a portfolio or a product description are not covered. A blog post on a matter of public interest can be, and then it is enough that a person reviewed it and stands behind it.

What the notice has to look like

Article 50(5) sets three conditions. The information must be

clear and distinguishable, so not hidden and not in the fine print,
present at the latest at the time of the first interaction or exposure, so visible as soon as the image is visible,
compliant with accessibility requirements.

Point three gets skipped a lot. A notice baked into a graphic as pixels reaches nobody using a screen reader. Write the disclosure as real text beside or below the image, not only inside it. Add a badge on the picture as well if you like, but the text underneath is the part that counts. Our post on building an accessible site covers how to do that properly.

Wording that works:

Made with AI
AI-generated image
Image: created with artificial intelligence

We would avoid hints like "digitally created" or "enhanced". They do not say what happened.

Quiet and still correct

Nothing in the law demands a badge stamped across the picture. What is demanded is that the information be clear and distinguishable and present at the latest at first exposure. That leaves room, and these four ways use it:

The caption. One line directly under the image, same typeface as the surrounding text, just smaller. The quietest option that reliably meets the requirement, because the line appears together with the image.
A small, consistent marker. A discreet "AI" in one corner, always in the same place, plus one sentence somewhere on the page explaining what it means. Kept up consistently, it costs almost no space per image.
One notice above a set of images. For a gallery where every image comes from an AI, a sentence above the set is enough, as long as it is visible together with the images. On a page where two of twenty images are generated, it is not.
The credits, for artistic and satirical work. Only there does Article 50(4) expressly allow the notice to sit where it does not hamper the work, for instance in accompanying information or the credits.

Two routes are not enough:

Metadata alone. The machine-readable marking is the provider's duty under paragraph 2. Your duty under paragraph 4 is a statement for people. C2PA in the file does not replace it.
A line in the privacy policy or the imprint alone. That is not present at first exposure of the image.

Who enforces it

In Germany the Federal Network Agency is the authority. Since 2 August 2026 it acts as market surveillance authority, point of contact and complaints body for the AI Act, under the national implementing act. Supervision is split: in already regulated product areas the established authorities keep their remit, and the agency coordinates and covers the rest. Other member states have named their own bodies.

Fines for breaching the transparency duties run to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. For a small company the realistic risk is less the fine than a competitor complaint.

What to do this week

Walk your existing pages. Which images came out of an AI? If you cannot remember, that already answers the next question.
Write down what was made with what. One line per image in a spreadsheet is enough. Anyone keeping that can answer a query in five minutes.
Disclose photorealistic AI images, visibly with the image and as real text.
Do not throw the metadata away. Plenty of image optimisers strip all metadata while resizing, which also deletes the machine-readable marking the provider put there. Check once what your tool does.
Only disclose in genuine doubt. The doubt rule belongs to the narrow strip between drawing and photograph. Where there is clearly no duty, no notice belongs either: labelling everything devalues the label where it counts.

What we do ourselves

There are no photorealistic AI images on primero.link. The example images in this post are drawn on purpose, partly for this reason: generating a convincing fake in order to explain when convincing fakes have to be labelled would be an odd move. By the rule above they carry no duty, so we put no notice under them either.

If you want a second pair of eyes on your own pages, drop us a line. A first read costs nothing.

Legal note

This post reflects the position on 6 August 2026 and is not legal advice. The AI Act is new, many detail questions have no settled interpretation and no case law yet. For a concrete project, talk to a lawyer.

Labelling AI images on your website: what is now required