AI labelling duty: what businesses have to label.

Four cases from everyday business

Not everything made with AI has to be labelled. In a company’s everyday work, these four cases matter most:

Whether the AI on your website is already labelled correctly is shown by the website check; it is free.

Who has to label what

The regulation distinguishes two roles. Providers develop an AI system or have it developed and place it on the market or put it into service under their own name or brand. Deployers use an AI system in their business, for example as a chatbot on their own website. If you use someone else’s system, you are usually a deployer; depending on development and provision, provider obligations can arise as well.

Under Article 50(5) of the AI Act, the notice must come at the latest at the first interaction or first exposure to the content, clearly recognisable and accessible.

What applies from when

The transition period came with Regulation (EU) 2026/1744 (Digital Omnibus on AI). According to the European Commission’s questions and answers, it only applies to marking by providers and only to older systems.

Chatbots and AI agents on your website

A chatbot or AI agent talks directly to people. Article 50(1) of the AI Act applies to that. In practice this means:

• Notice in the first sentence: the first message says that an AI is writing, for example “You are chatting with an AI”.
• Honest when asked: if someone asks “Are you a person?”, the answer is no.
• No human appearance: a name is allowed; a made-up photo with an employee’s CV is misleading.
• On the phone and by email too: the duty does not depend on the channel, but on a person interacting with the AI.

The first message could look like this:

Images, videos and voices

Under Article 3(60) of the AI Act, a deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic. Only for that does disclosure under Article 50(4) of the AI Act apply.

• Deepfake: an AI video in which your managing director says something they never said has to be disclosed.
• Art, satire, fiction: for evidently artistic or satirical works, a notice that does not spoil the enjoyment of the work is enough.
• Illustration without deception: a recognisably drawn or designed AI illustration does not resemble a real scene it pretends to show. It is usually not a deepfake. A voluntary note still builds trust.
• Photorealistic people: an AI image of a fictional person that looks like a real customer photo should be labelled.

Texts: only in one narrow case

For AI texts, Article 50(4) of the AI Act only applies when three conditions come together:

1. Published: the text is public, not just internal.
2. To inform the public: it is meant to inform, like a news item.
3. Public interest: it concerns a matter of public interest.

Even then, the duty does not apply if a person has reviewed the text for content and someone holds editorial responsibility.

Social media

Code of practice and fines

Checklist for your website

1. Chat and phone: does your chatbot or phone assistant say in the first sentence that it is an AI?
2. Images: are there photorealistic AI images of real or fictional people that look real? Then label them.
3. Videos and voices: are there AI videos or AI voices that imitate a real person? Then disclose them.
4. Texts: do you publish AI texts on matters of public interest without human review? Then label them or have people review them for content, with editorial responsibility for publication.
5. Records: write down who checks which content and how you label it.
6. For agencies and client projects: decide per project who implements the notice, who checks it after changes and who records it.

Your next step