AI labelling obligation: what companies must label since August 2026

Who is affected: provider or deployer?

The AI Act distinguishes two roles. Providers develop an AI system or place it on the market under their own name. Deployers use an AI system in a professional context. Most companies are deployers: they use a chatbot, an image tool or an AI assistant offered by someone else.

Your role determines which Art. 50 obligations apply to you. If, for example, you offer your own chatbot on your website under your own name, you may be a provider yourself. Purely private use is outside the AI Act.

The four cases of the labelling obligation

Art. 50 covers four situations:

• Interaction with AI (para. 1, providers): anyone talking to a chatbot or voice assistant must be able to tell it is AI, unless this is obvious.
• Machine-readable marking (para. 2, providers): systems that generate synthetic text, images, video or audio must mark their output as AI-generated in a machine-readable way, for example via watermarks or metadata. Assistive functions for standard editing that do not substantially alter the content are exempt.
• Emotion recognition and biometric categorisation (para. 3, deployers): the people concerned must be informed.
• Deepfakes and texts of public interest (para. 4, deployers): deepfakes must be disclosed. So must AI-generated texts that inform the public on matters of public interest, unless they have undergone human review and someone holds editorial responsibility.

For systems placed on the market before 2 August 2026, providers have until 2 December 2026 to implement machine-readable marking (Art. 111(4), inserted by Regulation (EU) 2026/1744). The other Art. 50 obligations apply without a transition period.

What is a deepfake, and what is not?

A deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places or events and would falsely appear to be authentic (Art. 3(60)). What matters is whether a person could mistake the content for a real recording.

A photorealistic AI image of a person in your advertising therefore has to be labelled. A clearly stylised illustration usually does not. For artistic, satirical or fictional works, a disclosure that does not hamper the display is sufficient. Where the classification is unclear, labelling is the safe choice.

AI texts: when is human review enough?

For texts, the obligation only applies if they are meant to inform the public on matters of public interest. That covers, for example, news or statements on social issues, not every product description.

Even then, the obligation does not apply if a human has reviewed the text and a person or organisation holds editorial responsibility. Record who approves which texts. That lets you show, if in doubt, that the exemption applies.

How to label correctly

The notice must be clear and distinguishable, appear at the latest at the first interaction or exposure, and be accessible (Art. 50(5)). In practice, the following works well:

• Chatbot: a clearly visible notice directly in the chat window, not just in the privacy policy.
• Images and videos: the notice directly in the image or video, for example as a badge. A caption alone is lost as soon as the content is shared.
• Files: an additional note in the file metadata.
• Social media: a notice in the image and a line in the post text. Many platforms also offer their own AI labels.

How we do it ourselves: all HabichtAI product videos and AI-generated images carry the notice “AI-assisted” or “Image AI-generated” directly in the image, plus an explanation in the caption and in the file metadata. For genuine screen recordings, we deliberately do not write “AI-generated” but “created with AI assistance”.

Checklist for your company

1. List all outward-facing AI applications: chatbots, image and video tools, text generators.
2. Clarify your role per application: are you a provider or a deployer?
3. Give chatbots and voice assistants a clear AI notice.
4. Set rules for images and videos: what counts as a deepfake, what does the notice look like?
5. Document the approval process for AI texts: who reviews, who holds editorial responsibility?
6. Train staff who create AI content (Art. 4, AI literacy).