AI Act: When must content generated by AI be labeled on a website?

From August 2026, the EU’s Artificial Intelligence Act (AI Act) will require that content generated or modified by AI that appears to be authentic be clearly labeled.

This obligation applies to content regardless of the publication channel. In this article, we focus in particular on images and other media content published on websites. In addition to images, the regulations also apply to videos and audio, as well as certain texts generated by artificial intelligence.

Artificial intelligence can be utilized in content in many ways, and not all uses require visible labeling. Nor is there any obligation to use the official EU icons. For website administrators, the key is knowing when content needs to be labeled and how to make the label clearly visible.

To help with this, we have developed a solution for WordPress sites that allows AI labels to be added to images and videos directly within the media library.

The AI Act requires content generated by artificial intelligence to be labeled

Article 50 of the EU’s AI Act concerns transparency. It stipulates that viewers must be informed if content that appears to be authentic has been generated or modified by artificial intelligence. These obligations will come into force on 2 August 2026, and in Finland, compliance will be monitored by Traficom, subject to certain exceptions. However, content published before 2 August 2026 does not need to be labeled retrospectively.

Responsibility for labeling is shared between the provider of the artificial intelligence system and the deployer. The provider must label the content it produces in a machine-readable format, for example using a watermark or metadata. The deployer – for example, the website owner – is responsible for ensuring that the labeling is visible to users. The machine-readable labeling added by the tool is therefore not sufficient on its own; the website must also display a label visible to the user, such as the EU AI icon.

For a more comprehensive overview of the AI Act, its risk categories, and implementation timeline, see our article here

Generated by AI or edited by AI?

An ‘AI-generated’ image is created entirely by artificial intelligence: human provides the prompt, but the image itself is generated entirely by AI and there are no human-made elements included. An image created entirely by AI must be labeled if it appears realistic. Clearly fictional or abstract content, such as a decorative background pattern, does not usually require labeling.

An ‘AI-modified’ image is based on a photograph taken or created by a human, which is then modified using artificial intelligence. If the AI is merely used to enhance an existing image – for example, by removing a minor distraction from the background or adjusting the color tones – no label is required. If AI is used to alter a human-taken or human-created image in such a way that it becomes new, realistic-looking content – for example, by changing the background or adding furniture to an empty room – the image is partially modified by AI and a label is required.

In practice, the line is drawn based on the extent to which artificial intelligence has contributed to the final result. An image generated entirely by AI is labeled as such when it looks authentic, while an image edited using AI is labeled as such when the editing creates new, authentic-looking content. Minor editing of an image alone does not require labeling, as this is something we do all the time using image-editing tools.

What does it mean when an image looks authentic?

Not all images created using artificial intelligence therefore need to be labeled. The decisive factor is whether the image appears genuine to the viewer. This type of content is known as a deepfake, and it is the responsibility of the publisher to label such content.

What is deepfake?

A deepfake is an image, audio clip or video generated or manipulated by artificial intelligence that looks or sounds genuine, even though it is not. It may resemble a real person, object, place, animal or event – or something that could exist in reality – but it may also depict a completely fictional, yet lifelike, person.

The term is best known in connection with scams and fake news. From a legal perspective, however, what matters is not whether the intention is to deceive, but rather that the viewer might be misled into believing the content to be genuine. According to the Commission’s guidelines, even a realistic, fictional character can constitute deepfakes, even if there is no one recognizable in the image.

A realistic image created using artificial intelligence, such as the cover image for this article, is therefore a deepfake, whereas nobody is likely to believe that the image of a dragon is real.

How is an AI image labeled?

The European Commission has published three official icons: a basic icon (AI) and icons for content generated by artificial intelligence (AI Generated) and content modified by artificial intelligence (AI Modified). The basic icon indicates the general use of AI, and explanatory text may be added alongside it.

The use of EU icons is voluntary, but where content is subject to the labeling requirement, labeling is not optional. Labeling may therefore also be carried out in some other clear manner. The Commission’s icons provide a ready-made and standardized way of doing this, and they can be downloaded from the Commission’s icons page.

The icon must be placed alongside the image in such a way that the viewer can clearly see it and recognize it no later than when they first see the image. The label must be large enough and stand out clearly from the image. A short text label may be added to accompany the icon, for example, ‘Image generated by artificial intelligence’. A mere mention in the terms of use, footer or image alt text is not sufficient, as the information must also be visible directly to the user.

The Commission also encourages consideration of accessibility. It is advisable to use simple language in text labels and to avoid jargon. Icons should also be given alternative text or an ARIA attribute – that is, a code annotation intended for screen readers – so that the information is also conveyed to screen reader users.

What about video, audio and text?

The same principles also apply to video and audio. Here too, a clear distinction is made between content generated entirely by artificial intelligence and content modified using artificial intelligence, although images are the most common form on most websites. A voice-over generated by AI or a video featuring a real person that has been edited using AI may require a disclosure. Background noise removal does not require a label, but AI-generated transcrip should be mentioned, as it is good practice to inform the reader, in the interests of transparency, that the text was not produced by a human.

In the case of texts, the transparency requirement applies to AI-generated texts of public interest that have not been reviewed by a human and for which no one bears editorial responsibility. In artistic, satirical or fictional works, it is sufficient to indicate the use of artificial intelligence in an appropriate manner that does not detract from the experience of the work.

Traficom provides further details on its website regarding when the use of artificial intelligence must be disclosed in relation to videos, audio and text. The page also contains guidelines for situations where a breach of the transparency obligation is suspected.

A WordPress solution for AI labeling of media

With the feature we have developed for WordPress websites, an AI labels can be selected for an image or video directly within the WordPress media library. The publisher does not need to add the tag to the image separately in an image editing software; instead, the site automatically displays it.

In the media gallery, select the appropriate label – AI, AI Generated or AI Modified – from the image or video details. There are several color options available for the labels to ensure they stand out against different image and video backgrounds. Once a label has been selected the label is displayed automatically on the image or video wherever that media is used on the site. The label also adapts to the screen size, ensuring it remains clear on both computers and mobile devices.

If your website features images and videos generated or edited using artificial intelligence, we can implement AI labeling on your WordPress site. Please get in touch and we’ll tell you more about the solution.