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

Updated: 13 October 2026

The short answer

Not all of them. Since 2 August 2026, EU rules require a visible label only for so-called deep fakes: AI images that resemble real people, places, objects or events and could be taken as authentic. An abstract or cartoon-style illustration for a blog post is normally not covered.

What the rules say and their status

The EU Artificial Intelligence Act has a transparency rule that has applied since 2 August 2026. It does not require a label on every AI image. You must disclose that an image, audio clip or video was generated or altered by AI only if it is a deep fake.

The detail comes from European Commission guidelines and a Code of Practice. The guidelines are not binding, and their formal adoption in all languages could not be confirmed by 13 October 2026. The Code is voluntary. Both still show how the rule is read.

Images generated or altered before 2 August 2026 need no label after the fact. Reworking an old image with AI after that date counts as a new alteration.

When an image is a deep fake

The guidelines set four conditions that must all be met. The image resembles something. That something exists or could plausibly exist. It is a person, object, place, entity or event. A viewer could wrongly take the image as authentic or truthful.

The person need not be real: it is enough that such a person could exist. Your intention makes no difference. Photorealism alone does not settle it. If children or older people are likely to see the image, it may be enough that they could be misled.

A realistic AI "photo" shown as your team, your customer or your salon is a clear case. A generic mood image is a weaker case, but not safe by default.

Examples: what is in and what is out

In scope: an AI image of a product in an advert or on packaging that makes it look more appealing or better than it really is. A realistic AI avatar presented as a real person.

Out of scope: a real product, such as a car, shown against an AI-generated background, as long as the advert does not mislead about the product itself. Minor edits to an existing photo: removing a passer-by, lighting, colour correction, cosmetic retouching. Scenes nobody would take as real, such as a dragon or talking mice in a cheese advert.

Not every stylised graphic is exempt. But a typical abstract or cartoon-style illustration does not pass itself off as a genuine photo, so it needs no label.

Who is responsible

The duty falls on whoever decides, professionally, to use AI and how. If your employee made the image, or a freelancer acting on your behalf and under your control, your business is responsible.

If an agency decided for itself whether and how to use AI, the agency has the duty and a client who merely commissioned the work does not. A client who asks for AI images, picks the tool or knowingly approves them has the duty too. So ask your agency whether it used AI.

A site that only publishes other people's images, such as stock photos, does not have this duty, though the Commission strongly encourages keeping existing labels. Private, non-professional use is not covered.

What the label should look like

The information must be clear and visible when someone first sees the image, for each image separately. A mention in the terms and conditions is not enough. Nor is the hidden mark the generator writes into the file: people must see the label without tools or clicks.

The rule itself prescribes no form. The Code of Practice puts the label on the image itself, for example in the top right corner, or in an overlay that appears to be on it. Showing it only on hover falls short. Whether a caption under the image is enough has not been decided, so a label on the image is safer.

You can use the free, optional EU "AI" icon. A label does not repair a misleading image: a product shown as better than it is still misleads the customer.

National differences and detection

Some countries add stricter rules, but for media or advertising, not for every website. Slovenian media label every item made with generative AI, at the beginning and the end. Croatian media and electronic publications label any AI-generated content. Romanian audiovisual media show an on-screen warning for the whole duration of realistic synthetic content. Latvia requires a label on paid election advertising where AI shows a person or event that does not match reality.

In Slovenia and Hungary an authority can already fine a missing label, although we found no report of a fine. In Poland this becomes possible on 28 October 2026. The duty itself applies everywhere now.

Automatic detection of AI images is unreliable. In a 2026 study, detectors tuned to recent generators flagged 39 to 63 per cent of genuine smartphone photos as artificial. So the decision is yours: keep a simple list of your AI images.

This describes the rules, not your particular case. Legal basis: Article 50(4) and (5) and Article 3(60) of Regulation (EU) 2024/1689.

Key points

Only deep fakes need a visible label: AI images that could be taken as authentic.
A real product on an AI background and minor retouching need no label. An AI image that flatters the product does.
Whoever decides to use AI is responsible: you, your agency or both, depending on who took the decision.
A label on the image, or an overlay on it, is safer than a caption alone. The generator's hidden mark is not enough.
Automatic detection of AI images is unreliable, so keep your own list.

Sources

This guide is for information only and is not legal advice.