AI disclosure texts to copy: chatbot, voicebot, deep fake and AI-generated text

Paragraph and role: 50(1), provider. If you use a supplier’s chat, you check as deployer that this text stays in place.

Placement: as the first message, before the first answer, plus a fixed label at the input field.

Paragraph and role: 50(1), provider. As deployer you check that the notice stays in the call script.

Placement: in the first sentence, before the question how you can help. Repeat the notice when the AI takes over the call again after a queue or transfer (point 37).

Paragraph and role: 50(1), provider, if an AI agent itself writes and sends emails or responds to replies. The agent says that it is AI and on whose behalf it acts (points 31 and 36). If an employee sends each draft personally, paragraph 1 does not apply (point 30).

Placement: at the top of the email, before the content, and in the sender name.

Paragraph and role: no obligation of your own as deployer under Article 50 if the image is not a deep fake. The provider of the image generator marks the image in a machine-readable way (50(2)). A visible caption is voluntary under Article 50 and helps your readers’ trust. Other law may still require a disclosure, for example in advertising where leaving it out would be misleading (Articles 6 and 7 of Directive 2005/29/EC; in the Netherlands Articles 6:193c and 6:193d of the Dutch Civil Code); see Article 50(6).

Placement: directly below or next to the image. If the image is realistic and could be taken for real, use text 3.5.

Paragraph and role: 50(4), first subparagraph, deployer. The tool’s watermark does not count: the label must be visible or audible without a tool (point 117).

Placement under the code of practice (section 2): for images and video a label in the top right corner, not overlapping other elements; for video at the start, where possible at intervals and after advertising breaks; for audio a short spoken notice at the start, repeated for long or live content. Use the EU ‘AI’ icon from the code of practice or an equivalent label.

If you work with an agency, freelancer or platform: agree in writing that the label is carried over and stays in place (points 12 and 14).

Older content: a deep fake you made before 2 August 2026 does not need to be labelled retroactively. If you still distribute it now, the guidelines do recommend a label (point 154).

Paragraph and role: 50(4), second subparagraph, deployer. Only for text you publish with the purpose of informing the public on matters of public interest. A product description or an email to a single customer is not covered.

Exception: the Act lifts the obligation if the text has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for the publication (Article 50(4)). The guidelines take this to mean that someone with knowledge of the subject reviews the content of the text before publication and that the person who holds editorial responsibility is publicly identified (points 133 to 140). Only checking spelling, a policy on paper or an AI checking an AI is not a review (point 135). We therefore recommend keeping evidence of the review for each publication.

Placement: above the text or with the headline, visible as soon as the reader starts. If only part was generated with AI, you may label only that part.

Older text: text you created before 2 August 2026 but publish only now is also labelled (point 154).

Paragraph and role: 50(3), deployer, for biometric categorisation that is not prohibited, such as estimating age (point 104). Combine the notice with your privacy notice (point 110). Also check whether the categorisation is high-risk: Annex III, point 1(b), covers biometric categorisation according to sensitive or protected attributes or characteristics (see chapter 5). The high-risk requirements then apply from 2 December 2027; for systems already placed on the market or put into service before that date, only after a significant change in their design (Article 111(2)). Also check whether a data protection impact assessment (DPIA) is needed (Article 35 GDPR).

Placement: a sign at every entrance to the area or a notice on screen before the start, at the latest at the first exposure (points 107 and 108). Inform children too, in language they understand (point 105).