AI Act: When Must You Label AI-Generated Content?

How the news changes for your business

The European Commission indicates that the transparency obligations concerned by Article 50 have applied since August 2, 2026. It has published guidelines to clarify their implementation. Thearticle 50 of the regulations remains the starting point for understanding the categories covered, the respective roles and the exceptions provided for. Check the current version before making a compliance decision.

This guide offers a method of inventory and editorial implementation. It does not constitute a legal qualification of your system. An onboarding chatbot, edited video, and copy-edited copy should be reviewed separately. For a sensitive case, prepare the factual elements and have the analysis validated by your counsel, your data protection officer or the competent person.

Distinguish between the supplier and the user company

Start by identifying your role for each tool. Using an off-the-shelf service is not the same as developing or marketing a system under your name. Certain obligations relate to suppliers, in particular technical marking mechanisms. Others concern the companies that deploy the systems and distribute certain content. The contract with the service provider can clarify operational responsibilities, but it is not enough to erase applicable legal obligations.

Create a file by use: tool, version, function, manager, audience and distribution channel. Ask the supplier what measures they put in place and what you must preserve when exporting. An image downloaded and then reworked may lose technical information. So avoid a blanket response like “the tool takes care of everything” without checking the actual path to the published content.

A chatbot must be recognizable as such

Article 50 provides for the information of people when they interact with certain AI systems, with the conditions and exceptions detailed in the text. For a business site, make the nature of the assistant understandable upon entering the conversation. A small mention buried in a legal page does not produce the same effect as an indication visible at the time of the exchange.

Example of wording to adapt: ​​“You are interacting with an automated assistant. For price or availability confirmation, our team will respond to your request. » This sentence must correspond to actual operation. Add a way to contact someone and explain useful boundaries to the visitor. Don’t give the chatbot a false professional history or a human role that it doesn’t have.

Review images, voices and videos

The regulation aims in particular at the disclosure of the artificial or manipulated nature of content constituting hyperfakes, according to its definition and the terms provided. It is therefore necessary to examine what is represented and the risk of confusion with an authentic scene. A stylized illustration of a process and a fake video showing a satisfied customer do not raise the same questions.

For your communication, adopt a simple editorial rule: photos presented as your creations must show your achievements. If you use a project simulation, say so clearly. Maintain source files and media origins. Even where a situation does not fall into a specific AI Act category, a misleading representation can undermine trust and raise other legal issues.

Treat texts according to their purpose

Article 50 distinguishes certain texts intended to inform the public on matters of public interest and provides in particular for conditions linked to human control and editorial responsibility. This provision should not be transformed into a universal obligation that is identical for all product descriptions. The mere presence of a rereading does not allow the conclusion that all obligations have disappeared, whatever the usage.

Document your production chain: sources consulted, intervention of the assistant, corrections and person responsible for publication. Serious proofreading checks facts, quotes, figures and conclusions. Our method for control an AI response provides an operational basis. Keeping track of these checks helps explain how the content was produced if an error is reported.

Prepare a decision grid by content

For each publication, ask four questions: is there interaction with an assistant, realistic representation of a person or event, information on a subject of public interest, or technical marking provided by the service? Write down the answers without asking the AI ​​to certify compliance itself. Ambiguous cases should be reread with the applicable text and guidelines.

Fictional example: a store uses three functions. His assistant responds to schedules; an illustration explains in-store collection; a synthetic video shows an alleged client. The grid leads to three distinct treatments. The first must be clearly presented to the visitor, the second documented according to its nature, and the third must never be used as a real testimony. This separation avoids decisions made solely based on file format.

Put the mentions in the right place

A statement must be legible when it helps the public understand what they are seeing or using. Test it on the phone, in the chatbot window and on the platforms where you republish your content. Image crops, captions, and previews can change what remains visible. Also check that the necessary information survives the export from your creation tool.

Entrust this control to a person identified in your publication process. It verifies the nature of the content, the location of the mention and the retention of relevant technical information. Prepare a correction procedure for old publications. Our article on misuses of AI shows why the final check must cover the consequences of the content, not just its writing quality.

Deploy a lightweight procedure within the team

Start with an inventory of existing uses: site, social networks, emails, commercial supports and internal tools. Choose five representative examples to form the team. For each, explain what needs to be verified, who decides, and where to keep proof of validation. A procedure applied to business examples is easier to remember than a general presentation of the regulation.

Plan a review when changing tools or functionalities. An assistant who prepares drafts can then suggest sending or video generation; its old framework does not necessarily cover these new functions. Link this review to your AI implementation plan. The rules must follow actual uses, applicable texts and the information given by your suppliers.

Frequently asked questions

Does any spelling correction require an AI mention?

A general obligation should not be inferred from each use of an assistant. Examine the type of use, the categories of section 50 and its exceptions.

Is a mention enough to make false testimony acceptable?

No. Reporting an artificial generation does not turn an invented experience into customer proof. Present your demonstrations and simulations for what they are.

Can we delegate all control to the supplier?

Ask for their guarantees and instructions, then check the obligations related to your own role. Your publication and your interface remain to be examined in their real context.

What to do if the team is hesitant about content?

Maintain context, tool, version, and intended destination. Submit this set to a competent person before distribution, rather than choosing a random mention.