Labelling AI-generated content: symbol, placement, and exceptions

A deployer is anyone who uses an AI system under their own authority in the course of a professional activity (Article 3(4)). Deployers have a labelling obligation in three cases:

• for deepfakes (Article 50(4)),

• for published texts on matters of public interest (Article 50(4)), and

• when using emotion recognition or biometric categorisation (Article 50(3)). This applies to you only if you use such systems, which is not the case for the IONOS products listed below.

You are a deployer when you use the following IONOS products and features:

• AI App & Site Builder
• AI Receptionist
• AI Chat Assistant
• AI assistance in MyWebsite Now

Under Article 99(4)(g) of the AI Act, a breach of the transparency obligations can result in fines of up to €15 million or 3% of total worldwide annual turnover for the preceding financial year. Whichever of the two amounts is higher applies; for small and medium-sized enterprises, whichever of the two is lower applies.

Deepfakes in images, videos, and audio files

If you publish artificially generated or manipulated image, audio, or video content that resembles real people, objects, places, entities, or events and could falsely appear to be authentic, you must disclose that the content has been artificially generated or manipulated.

What matters is not whether a specific, identifiable person is shown. It is enough that the person, place, or object depicted could exist in reality. A photorealistic image of an entirely invented member of staff is therefore a deepfake. Depictions that are recognisably impossible are not covered, and neither are cartoons, drawings, and illustrations that have no potential to deceive.

Photorealism is a strong indicator, but it is not decisive on its own. What matters is whether the content is capable of deceiving people about its authenticity in the foreseeable context of use. An intention to deceive is not required – the assessment is made objectively.

Labelling is required in the following examples:

• an AI-manipulated image of a scene showing two real professional footballers in front of a building that resembles a football stadium
• an AI-generated video of a person who resembles a politician giving a speech to an audience
• AI-generated audio in which the voices of a podcast’s regular presenters are cloned
• a video featuring a realistic synthetic avatar of the management team congratulating employees on the annual results
• an AI-generated product image in advertising or on packaging that could mislead people about the product’s actual appearance, properties, or use – for example because the product looks more attractive, or simply different, than it does in reality

Minor image edits

AI-assisted colour correction, extending a background, replacing a background for purely aesthetic reasons, arranging existing products, or scaling images in product advertising and packaging design do not usually change the perception of authenticity significantly, and therefore do not trigger a labelling obligation.

Journalistic images

With journalistic images, however, substantial AI editing of background details that goes beyond customary technical and editorial practice can indeed affect the perception of authenticity.

Art and satire

Reduced obligations apply to evidently artistic, creative, satirical, fictional, or comparable works. In those cases, the labelling only has to be applied in a way that does not hamper the display or enjoyment of the work.

AI-generated texts on matters of public interest

If you publish AI-generated or AI-modified text in order to inform the public on matters of public interest, you must disclose that the text has been artificially generated or manipulated.

The obligation only applies if all three criteria are met together:

• The text is published. It is accessible to an indeterminate, larger number of readers who are not connected to one another – whether free of charge or paid for, such as by subscription. Individual private or business correspondence and texts used internally within an organisation, such as intranet posts, are not published.
• The text informs the public. It conveys knowledge, opinions, or facts. Short texts that convey none of these are not covered.
• It concerns a matter of public interest. This means topics that are relevant to society as a whole – locally, nationally, at European level, or internationally – and that merit public debate or scrutiny. The guidelines cite politics, justice, environmental protection, and consumer safety, as well as economic developments that may be the subject of public debate. What counts as public interest can change over time and depending on the context.

Product descriptions, category texts, landing pages, advertisements, replies to customer reviews, and social media posts about your own offering are generally not texts on matters of public interest. There is no labelling obligation under Article 50 in these cases.

The exception: human review and editorial responsibility

The labelling obligation for text does not apply if two conditions are met together:

• the text has undergone a process of human review or editorial control, and
• a natural or legal person holds editorial responsibility for the publication.

The requirements for that review are stricter than they first sound:

• A substantive review by a person with relevant knowledge and professional judgement of the subject is required. Checking the facts is the bare minimum.
• Superficial, purely formal, or procedural checks are not sufficient – for example a spelling and grammar check, merely having an editorial policy in place, automated checking steps, or a cursory sign-off without engaging with the content.
• Any substantial AI intervention after sign-off cancels the exception. If you have an approved text rewritten, extended, or reworded by an AI tool afterwards, you must label it again. Watch out for this with automated SEO and text-optimisation features.

Editorial responsibility means bearing the ultimate legal responsibility for the publication, including for the review itself. That can be an individual, an editorial team, or the publishing company. Make this point of contact publicly available and easy to find: an identity and a way of getting in touch, where a job title is also sufficient. Suitable places are your legal notice, your terms of use, or other legal information on your website.

The simplest option is to use the official EU symbols. They show the acronym AI, are free of charge, and are available in three variants:

Images and videos

• Images: place the symbol where no other elements overlap it – in the top right-hand corner, for example, or directly in the caption. Also add an accessible description to the image’s alt text.
• Videos: display the symbol at the start of the video. Where possible, also show it at regular intervals or for the whole video, and at least after breaks such as advertisements. The reason: viewers join at different points, and videos get turned into screenshots and clips. Showing the symbol continuously, or at least for the duration of the AI-generated section, is recommended.
• Internal, closed use, such as staff training: if a deepfake is used solely within a closed internal professional context, the disclosure can be placed in the user interface, in the physical environment, or through another suitable medium. It must still be easy to access.
• For visual deepfakes, audible disclosures may only be used in addition. They must always be accompanied by visual labelling.
• Social media thumbnails: place the symbol as an overlay in the image or video. Placing it in the caption alone is not sufficient.

Audio deepfakes

• Put a short spoken disclosure in clear, plain language at the start of the deepfake itself – in the language of the content, or in English. Where it fits, add whether the content is AI-generated or AI-modified.
• Another acoustic form is also permitted, such as a recognisable audio cue, or earcon. This does require you to accompany it with explanation – notices or repeated clarifications, for example – for as long as there is no uniform EU solution.
• For long or live recordings, repeat the disclosure at regular intervals, and at least after breaks such as advertisements. There is a practical reason for this too: long recordings get turned into excerpts.
• If a screen is available – in a car or on a smartphone, for example – you must provide visual labelling in addition to the spoken announcement.

Texts on matters of public interest

• Place the labelling above or at the start of the text, near the headline or in the colophon at the beginning of the text. What matters is that the placement is clear, consistent, and distinguishable for readers.
• If only part of the text is AI-generated or AI-modified, you may label just that part.
• For very short outputs, such as individual words or short phrases, where labelling directly on the text would impair readability or usability, the labelling remains mandatory. It may then be moved into the user interface: as an indicator immediately next to the output, or as a notice at the start of the session or on first interaction.

If your website offers an AI chatbot or AI assistant that people interact with directly, users must be able to recognise that they are communicating with an AI and not with a person (Article 50(1)) – unless that is obvious anyway. A simple notice at the start of the chat is usually sufficient. The provider of the AI system must make this function available to you.