A company is launching a new product. The marketing team uses AI to prepare the advertising visuals, uses a digital twin of the CEO in a video instead of filming, and generates an article about the product’s benefits. On the website, an AI assistant answers customers, helping them choose and dealing with questions about their orders.
One campaign, several uses of AI — and different obligations. What does the company have to label? Is it enough for someone to approve the article before it is published? And what does the AI tool provider or the advertising agency have to take care of?
If the CEO personally recommends the product in a video without having filmed anything, or a customer on the website assumes they are talking to a person, they may give the message more weight than they would if they knew the truth, and that can affect their decision to buy. Informed decision-making and a lower risk of deception are exactly what the transparency rules under Article 50 of the AI Act, which have applied since 2 August 2026, are meant to support.
Article 50 sets out four cases (see below) in which the use of AI must be disclosed. Their practical limits are explained in the Commission guidelines; practices for marking content are further developed in a voluntary code of practice.
1. It must be clear that AI is answering
When people interact directly with AI, its involvement must be obvious at the latest at the first interaction. This applies, for example, to a chat on an online shop where a visitor asks which product to choose.
The obligation applies: an AI assistant conducts the conversation and recommends a product. A practical solution is a notice right in the chat window: “You are chatting with an AI assistant.” A mention hidden in the terms and conditions is not enough. An exception is where the involvement of AI is obvious from the circumstances; the label “assistant” alone, however, cannot be relied on.
The obligation does not apply: AI merely suggests a reply to a support agent, who checks it and sends it themselves. The customer is therefore communicating with a person, not directly with AI.
The notice must be ensured by the provider of the system. That may also be a company that has a system developed for it, or that adds its own data to a third-party solution and puts it into service under its own brand. When working with a supplier, it is therefore necessary to clarify the roles and agree in the contract who sets up the notice, maintains it and makes sure it keeps working when the system changes.
2. Biometric categorisation and emotion recognition require a notice
A company is presenting a product at an event. A camera at the entrance estimates visitors’ age groups. The company also offers members of the public the chance to test the product voluntarily, while AI estimates their emotions from their facial expressions. The company must inform the people concerned about both the biometric categorisation and the emotion recognition.
The obligation applies: AI sorts visitors into age groups based on their faces, or estimates, for example, surprise or disappointment from their facial expressions.
The obligation does not apply: visitors state their age themselves when registering and give their impressions in a questionnaire, whose written answers AI sorts into positive and negative. This is not an analysis of biometric data.
The notice must be available before people enter the area where the analysis takes place, and it must make clear that AI is estimating age and recognising emotions. General information about a CCTV system is not enough. The obligation also applies where a recording is analysed later.
Informing people does not, however, make the analysis lawful. Above all, the requirements of the GDPR must be considered. In the workplace and in educational institutions, emotion recognition is prohibited in principle, with exceptions for medical or safety reasons.
3. Videos and articles may need a visible label
Back to the campaign. For images, audio and video, what matters is whether the content is a deepfake, i.e. content generated or manipulated by AI that resembles real persons, objects, places, entities or events and would falsely appear to be authentic. For text, what matters is whether it is published with the purpose of informing the public on matters of public interest.
Video and advertising visuals
Label: a realistic video in which a digital twin of the CEO presents the product, even though the CEO filmed nothing.
Generally no label needed: fine-tuning light and colour, or an AI background behind a real product, provided it does not distort the product’s appearance, properties or use.
A generated image of a box is not a deepfake merely because it was not photographed. What matters is the overall impression: does the content create an appearance of authenticity or not? A visualisation that looks like real evidence of the product’s effects may require a label.
An article on the company website
A label may be required: AI prepares a text about the product’s health effects. This is informing the public on a matter of public interest, even though it also supports sales.
Generally no label needed: a description of the packaging, number of tablets, flavour or storage of a food supplement, with no claims about health effects, safety or sustainability.
For text, there is also an exception: the obligation does not apply if the text has undergone genuine human review or editorial control and a specific person or company holds editorial responsibility for its publication. Review means a substantive assessment of the content and fact-checking by a person with the authority to change or reject the text — not proofreading, and not a check by another AI.
4. The origin of content must also be detectable technically
The marketing team uses an image generator. The provider of the tool (but not the organisation using it) must ensure that the outputs carry a machine-readable marking showing that they were generated or manipulated by AI. For this purpose, the code of practice mainly envisages a combination of digitally signed metadata in the file and invisible watermarks embedded directly in the content.
The obligation applies: the tool creates an advertising image, even if it shows an obviously fictitious scene that the company does not have to label visibly as a deepfake.
The obligation does not apply: the tool only corrects spelling or performs standard editing without substantially changing the content or its meaning.
Technical marking is for machine detection and cannot be seen by the human eye. For a deepfake, it therefore does not replace the notice that the content was generated or manipulated by AI.
What should the label look like?
The form of the label can be adapted to the website or campaign, but the AI Act sets out several requirements. The label must be clear and distinguishable, so it must not get lost in small print or be hidden in a caption nobody clicks open. Customers must learn about the use of AI at the latest when they first encounter the content. It could be, for example, a caption saying “Video created with AI” or “Voice modified with AI”.
The Commission has prepared three icon variants for labelling deepfakes and texts for which the AI Act requires disclosure of the use of AI. The icons are available on the Commission’s website and their use is voluntary.
 |
Content generated or modified with AI. Details can be added in a caption or made available on click. |
 |
Content fully generated by AI, with no human creative elements or editorial input other than prompting. |
 |
Content originally created by a person and then partially modified with AI, for example a photo with a replaced face. |
