AI content labelling: making the EU icons stick

Start at the end of the chain, not the beginning

Most writing on this subject starts with the regulation and works forward. If you sell through other people, it is more useful to start at the other end: with the screen a customer is actually looking at. That screen is often not yours. It belongs to a dealer, a retailer or a marketplace, and the image on it arrived there by being downloaded, resized, re-uploaded and sometimes exported in a different format.

Every one of those steps is a chance for a label to fall off. So before you decide which icon to use, decide where the label is going to live. Get that wrong and the icon is decoration.

What Article 50 actually asks

The transparency obligations in Article 50 of the AI Act have applied since 2 August 2026. The principle is simple enough: people should be able to tell when they are looking at content that was created or altered by AI.

Two kinds of content fall under the disclosure duty in paragraph 4:

• •Deepfakes: AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would appear authentic.
• •AI-generated text published to inform the public on matters of public interest, where nobody has taken editorial responsibility for it.

There are carve-outs. Content that is evidently artistic, creative, satirical or fictional carries a lighter duty. For text, the duty falls away where there is human editorial review and someone takes responsibility for it. And there is an exception for legally authorised law enforcement.

Whether a particular image in your catalogue sits inside or outside that definition is a legal judgement about your business, and not one we can make for you. We are not lawyers and this is not legal advice. Put the question to your own counsel. What you can do without waiting for the answer is record, for every image, whether and how AI was used, so the question is answerable at all.

One assumption worth dropping early: using somebody else's AI tool does not hand the obligation to them. The regulation places duties on the provider of the AI system and on the party deploying it and publishing the output. If you publish, that second party is you. For what you may and may not do commercially with an AI-generated image, see our piece on AI-generated images and commercial use.

The three icons

As part of the Code of Practice on marking and labelling AI-generated content, the European Commission has released three pictograms. They are free to use, in SVG and PNG, in black, white and a 50 percent transparent version.

• •Base pictogram. For when AI was involved in producing deepfake content or published text, or when you pair it with your own text label or a second layer of information.
• •AI Generated. For content produced entirely by AI, with no human-made elements. Writing the prompt does not count as the human part.
• •AI Modified. For existing human-made material that has been partly altered with AI.

These are the icons we use. The files and the usage rules are on the European Commission page on the EU icons for labelling AI-generated content.

Worth being precise about: using the icons is voluntary. The disclosure duty in Article 50 is not. An icon does not make you compliant on its own, and leaving it off does not release you from being clear.

The display rules, and the one that decides everything

The Code of Practice is fairly concrete about presentation. The main points:

• •Clear at first exposure. The pictogram has to be perceivable and distinguishable no later than the first time someone encounters the content. Not behind a click, not in the terms.
• •Embedded in the content itself, so it stays in place when the content is shared or downloaded.
• •Nothing layered over it. No platform badge, overlay or watermark on top of the pictogram.
• •Paired with plain language. The Commission's own research found the base pictogram works better alongside a text label. Keep that label free of jargon.
• •Accessible. Readable by assistive technology through alt text or ARIA labels, at a clearly visible size, and on screen long enough for people who need more time to process it.
• •A second layer is allowed. A tooltip or pop-up with more detail is fine, as long as assistive technology can reach it and the pictogram signals there is more behind it.

Read the second one again. “Embedded in the content itself” reads like a styling preference, and it is the rule that quietly determines your entire approach. It is the difference between a label that reaches your retailers and one that does not.

Two ways brands lose the label

• •It lives in the template. A caption reading AI-generated sits in the markup of your own product page. Your dealer downloads the file, not the page.
• •It lives in metadata only. That feels tidy, and it is fragile. Plenty of image editors and content systems strip metadata on resize or format conversion, with nobody noticing. Do not assume a marker you attach is still in the file at the end of the chain.

Both failures have the same shape: the label was attached to the context rather than to the image. The fix is to put it where the image goes.

And make it an agreement rather than an assumption. Whoever publishes, labels. That belongs in the brand guidelines you hand your channel, not in a one-off email to a dealer marketing manager who will have changed jobs by spring.

How we built it into LensLab

Since 9 November 2026, labelling sits in the platform. When a customer saves a delivered visual, they choose whether to put an EU AI content label on it, and which of the three: AI, AI Generated or AI Modified.

The icon goes into the bottom right of the image itself, rather than being shown only inside the platform. That means it stays visible after download, at the dealer, and everywhere the image travels afterwards. Closing exactly the failure described above was the point of building it that way.

Customers pick the style too: black or white, each with a transparent variant, so the icon stays readable on a light or a dark image.

Labelling is a choice rather than a default. It is off unless the customer switches it on for that delivery. Two limits worth stating plainly: it sits on the review step for delivered visuals, so not in every corner of the platform, and videos do not get a label. The label we apply is a visible icon, meant for the person looking at the image.

Around that sits the way we work. Every Visual Studio Pro delivery passes a creator from our own team before it reaches the customer. Human review does not answer your labelling question, but it does mean somebody saw what was made. How that keeps AI from inventing details your product does not have is in accurate AI product photos.

Four things to put in place this quarter

• •Record per image whether and how AI was used, in the system where you manage your images. Without it, every later judgement is guesswork.
• •Decide where the label lives, and choose the image over the page.
• •Write the rule into your brand guidelines: whoever publishes, labels.
• •Take the scope question to your own legal counsel, with your real catalogue in front of them.

The first of those is the one people underestimate, and it is the practical case for a properly run image library. We made that argument at length in our piece on the last mile of digital asset management.

In short

The rules apply. The icons exist, they are free, and they are well thought through. The hard part is not the pictogram. It is working out which of your images are covered, and then making sure the label survives the trip to everyone who sells your products. Start with the record per image, put the label in the image, agree the rule with your channel, and take the edge cases to your lawyer.

Curious how brands manage their imagery centrally and get it to their dealers? Have a look at our use cases or what LensLab costs.