Do I have to label AI-generated images on my website?

From 2 August 2026, yes, if the content is published and resembles real people, objects or scenes. Article 50 of the EU AI Act requires that it is recognisable to the viewer that an image, video or audio was artificially created or altered. Purely private or purely abstract-decorative use generally falls outside the duty - when in doubt, label it.

Does the EU AI Act prescribe a specific wording or icon?

No. The EU AI Act requires the disclosure to be clear and recognisable, but makes no requirement about wording, symbol or design. There is no compulsory label. The EU Commission's voluntary Code of Practice offers building blocks, but participation is explicitly optional. So you can decide how the notice looks - a discreet info symbol with an explanatory text is enough.

May I phrase the AI notice positively instead of writing 'AI-generated'?

Yes, as long as the notice stays truthful and recognisable. You may set the mandatory notice in an honest context - such as protecting customer data - and use softer terms like 'created with the help of generative AI'. What matters is that the use of AI is clear to the viewer and that the stated reason is true.

Does this also apply when the product is real and only the image was made with AI?

Yes. The EU AI Act looks at how the image came about, not at whether the depicted object is real. Even if your product genuinely exists and was merely visualised with generative AI, it counts as artificially created content and must be labelled. This is exactly where the positive frame pays off, because it explains why you work this way.

Labelling AI content on your website: turn the obligation into a trust signal

An AI label has never turned anyone into a customer. I have never looked at a website and thought: impressive, they use AI. But the label doesn’t have to put anyone off either. That is what this piece is about.

From 2 August 2026, the labelling duty for AI-generated content under Article 50 of the EU AI Act applies. Anyone who publishes AI-generated images, videos or audio has to label them so that it is recognisable to the viewer that they were artificially created. The reflex in most guides is: put a clearly visible “AI-generated” underneath, ideally in a warning colour. That is correct and still a wasted moment. Because the wording and the design of that notice are not prescribed by law at all.

One clarification up front, because it lands wrong otherwise: I do not think the labelling duty is excessive. Transparency about what is real and what is machine-made is right, and it matters more the better the tools get. My point is not to do less than the law requires. It is to implement what is mandatory anyway in a smarter way.

What Article 50 actually requires - and what it doesn’t

Article 50(4) of the EU AI Act addresses deployers, meaning everyone who publishes AI-generated content. The core duty: whoever publishes image, audio or video content that resembles real people, objects or events and was artificially created or altered must disclose that it is artificially generated or manipulated. The disclosure must be clear and recognisable.

What it explicitly does not say is just as important:

No prescribed wording. The law names no mandatory phrase. Neither “AI-generated” nor any other fixed text is required.
No compulsory icon, no standard label. There is no official symbol you must use. The EU Commission is developing a Code of Practice on labelling through the AI Office, which offers building blocks and recommendations - but taking part in it is voluntary. The duty to label exists; the concrete form is free.
No design requirement. Whether a discreet notice, a small info symbol that explains on tap, or a line in a caption: as long as it is recognisable to the viewer, the form is up to you.

There is a technical layer alongside this: the providers of the AI tools must mark their outputs in machine-readable form, for example via embedded metadata under the C2PA standard. The tools handle that, not you - you should simply not strip such markings out on purpose. Your job as a deployer is the visible disclosure. And that is open in its design.

Why a bare “AI-generated” is a weak label

Picture two notices on the same image.

The first: “AI-generated.” Two words, factually correct, duty met. What stays with the viewer is: this is not real. In the worst case it plants a small doubt that was not there before.

The second: “To protect our customers’ privacy, we show a depiction created with the help of generative AI.” One sentence, duty likewise met, the use of AI named clearly. What stays with the viewer first is: they protect their customers. The information “made with AI” is there, it is recognisable, but it does not sit as a cold tag at the front - it is embedded in a reason that makes sense.

Both notices meet the same legal duty. The second one also works for you. That is the whole trick: not to label less, but to frame the label.

The honest positive frame

The idea behind it is simple. A mandatory statement has to be true and recognisable - but it does not have to sound like a product recall. You may embed the same notice in the reason you actually work with AI. A few patterns that fit depending on context:

Data protection: “To protect our customers’ privacy, we use digitally created depictions.”
Business secrets: “To safeguard our clients’ trade secrets, we show images rendered with the help of generative AI.”
Consistency and quality: “For a uniform, high-quality presentation, we partly use computer-generated visuals.”

Two subtleties from practice. First, word choice: “rendered with the help of generative AI” or “digitally created” carries the same information as “AI-generated” without the negative tinge. Second, placement: the notice has to stay recognisable, but it does not have to be the loudest element on the page. A discreet info symbol on the image that reveals the explanatory text on tap is a good compromise between recognisability and calm design.

You have a live example right in front of you: the cover image of this article is AI-generated. In the top right sits a small info symbol. Tap it and a sentence explains that I create cover images with the help of artificial intelligence to give the articles visual value. No warning sign, no red bar - and still clearly recognisable. That is exactly what I mean.

The line: honest yes, misleading no

Now the part that matters as much as the rest, because without it the whole thing tips over. The positive frame is not a cloak.

Two rules draw the line. First: the reason you state must be true. If you write that you protect customer data, then that has to be the real reason. An invented noble claim would not only be unconvincing, it would quickly drift into misleading advertising and become legally risky. Second: the use of AI must stay recognisable. You may embed it calmly and positively, but not hide it so that no one can find it any more. Recognisability is the hard requirement of the law, and for genuine deepfakes - content that looks deceptively real and creates the impression of authenticity - it applies especially strictly, without the editorial exception that exists for plain text.

That also makes clear what the positive frame is not for: anyone who builds an actual fake, has a person say or do things that never happened, invents an event, must label it unmistakably. There I am fully on the side of the law. Concealing it would not be clever, it would be dangerous.

When the product is real and only the image comes from AI

The most interesting grey area is cases where nothing is faked and labelling is still required. An example from my work, deliberately anonymised.

A client could not provide usable photos of their product - time, equipment, circumstances, the usual. What they had were the parts: the real packaging material and the genuine label. From those I visualised the finished product and placed it in a fitting scene. The product exists exactly like that, it is genuinely sold, the label is real. I assembled it on the computer - with generative AI instead of hours of cut-out and retouching work in a classic image editor.

And this is where it gets interesting: I could have laboriously built the same image in Photoshop, and hardly anyone would think to demand a “manually composited” label. Just because a more efficient tool is involved, the result is suddenly meant to carry a negatively loaded stamp. The EU AI Act consistently looks at how the image came about, not at whether the product is real - so it must be labelled. Fair enough. But in exactly these cases the positive frame is not spin, it is the more honest description: the product is real, the rendering is digital, and the reason is legitimate. You can say precisely that.

How to go about it in practice

A short sequence to clear this off your website cleanly:

Take stock. Which published images, videos or audio are fully or partly AI-generated? Explicitly exclude real photos, logos and customer shots.
Clarify the need to label. Does the content resemble real people, objects or scenes? Then it needs labelling from 2 August 2026. Purely abstract-decorative is usually uncritical - when in doubt, label it.
Set the frame. What is your real reason for using AI - data protection, business secrets, consistency? Cast that reason into a short, true notice.
Choose the form. Discreet info symbol with an explanation, notice text or caption. It has to be recognisable, it does not have to be loud.
Treat deepfakes separately. Anything that looks deceptively real and could feign authenticity gets a clear, unmistakable label - no experiments here.
Apply it consistently. The notice belongs everywhere it applies: gallery, detail view, lightbox, cover image.

Conclusion

The labelling duty is coming, and it is fine. The real question is not whether you label, but how. A bare “AI-generated” meets the duty and wastes the moment. The same notice, embedded in an honest reason and in a calm element you design yourself, meets the same duty and builds trust instead of undermining it. The law explicitly leaves you this room. Use it - honestly, recognisably and without the warning-sign aesthetic.

Note: this piece reflects my assessment as a practitioner and is not legal advice. For a binding evaluation of a specific case - especially with deepfakes or advertising claims - consult a law firm specialising in IT and advertising law.