How to label AI content correctly: A practical guide for websites and social media

1. What Article 50 really requires—and what it does not

Article 50 does not impose a blanket labelling obligation for “anything involving AI”. What matters are the context, the type of content and whether deception is possible. Legislators want to prevent people from being unknowingly deceived by AI—not require a notice for every spelling correction.

The article distinguishes four situations:

• Direct interaction with AI (e.g. chatbots)
• Machine-readable marking of synthetic content (technical obligation for providers)
• Emotion recognition and biometric categorisation
• Deepfakes and certain AI-generated texts

For most SMEs, the first and fourth cases are particularly relevant. We will also briefly classify the other two.

2. Provider or deployer? Your role makes the difference

Before labelling anything, clarify your role. The AI Act clearly distinguishes between providers (the developers of AI systems) and deployers (companies that use these systems).

• Providers —such as OpenAI, Google or Meta—must ensure that their systems mark synthetic content in a machine-readable format, generally through watermarks or metadata. This technical obligation lies with the manufacturer, not with you.
• Deployers —which includes the vast majority of companies—are responsible for the visible labelling provided to their users, meaning the notice that website visitors or followers actually see.

In practice, this means that if you publish an AI-generated image from a common tool on your website, do not assume that the provider’s watermark is “enough”. As the deployer, you are responsible for labelling that people can recognise.

3. The four labelling cases in detail

3.1 Chatbots and AI assistants

Users must be able to recognise that they are talking to a machine rather than a person. The notice should appear at the start of the interaction, before the first exchange.

Example wording: “You are chatting with an AI assistant. For personal enquiries, we will be happy to connect you with a member of our team.”

3.2 Deepfakes (image, audio, video)

The labelling obligation for media applies only to deepfakes —that is, artificially generated or manipulated content that could appear authentic. A deceptively real “executive board video” or an invented photorealistic product image falls into this category. Simple colour correction or cropping, by contrast, does not turn a photo into a deepfake.

3.3 AI-generated texts on matters of public interest

Texts that inform the public about matters of public interest must be disclosed as AI-generated—unless a person has reviewed the content and assumes responsibility for it (more on this shortly). This obligation typically does not apply to purely promotional or internal texts.

3.4 Emotion recognition and biometric categorisation

If you use systems that recognise emotions or categorise people biometrically, you must inform the individuals concerned. This case is uncommon for most SMEs in a marketing and website context—but if it applies, you should have it assessed separately.

4. Important exceptions—so you do not over-label

Article 50 takes a practical approach. You should know these exceptions:

• Obviousness in the case of chatbots: If it is obvious to an averagely informed and attentive person that they are speaking with AI, an explicit notice is not required. Use this exception cautiously—it can be legally difficult to assess in individual cases.
• Editorial control of texts: If an AI-generated text has undergone a substantive human review before publication and a natural or legal person assumes editorial responsibility, the disclosure obligation may not apply.
• Obvious fiction: An AI image of a dragon above the Bolzano skyline is so clearly impossible that nobody would be deceived. Where deception is impossible, deepfake labelling is not required. For artistic, satirical or fictional works, a label that does not impair the work is sufficient.
• Existing content: There is no retroactive labelling obligation for synthetic content generated and published before 2 August 2026. Retrospective labelling is not required.

5. How to label content in practice

Placement

The label must be clear, understandable and timely —meaning where and when the user encounters the content: in the caption for an image, as an introductory notice for a chatbot, or in or directly alongside a social media post.

EU icons or your own label?

You can use standardised labels or your own brand-compliant seal. Important: an icon alone does not prove compliance—as the deployer, you remain responsible for correct labelling. Your own seal must be just as clearly recognisable and correctly placed as a standard icon.

Do not forget accessibility

Notices placed directly on images must be legible and offer sufficient contrast. Also include the information in the alt text —that is, state that the image content is AI-generated or AI-modified. This also meets the requirement for perceptible labelling for people with visual impairments.

Wording you can use

• Image caption: “This image was created using artificial intelligence.”
• Partially edited image: “AI-modified.”
• Social media post: “Notice: This post contains AI-generated content.”
• Alt text: “AI-generated depiction of … (created using AI).”
• Video/audio deepfake: A visible notice in the image and in the description: “Artificially generated content.”

6. Your next steps

Labelling is not an end in itself or a bureaucratic monster—when used correctly, it is a signal of trust. The most efficient approach is to proceed in this order:

1. Take inventory: Where do you use AI—in chatbots, image generators, text or voiceovers?
2. Classify: Which of the four cases applies, and is there an exception?
3. Assign responsibility: Who reviews AI-generated texts and assumes editorial responsibility?
4. Standardise: Define fixed wording and placements in an internal policy so that labelling is not left to chance.

At Nuviax, we provide pragmatic support with precisely this assessment—from taking inventory of AI and clarifying roles to creating a ready-to-use AI policy tailored to your actual use of AI.