EU AI Act · Article 50 · Luxembourg

AI Act transparency: what you have to label, and by when

Remember the scramble when GDPR landed: the emails, the pop-ups, the "we've updated our privacy policy"? Article 50 is the AI version, and here is the reassuring part: it asks far less of you than the noise suggests. Since 2 August 2026, if people talk to your AI or see something it made, you usually have to say so. That is most of it. This is the plain-language reference for a small Luxembourg business: the four things it covers, what a real label looks like, and what you can safely ignore. No 60-page memo. No fear-selling.

Take the 2-minute readiness check Or book a call

The one-line version

If someone could be fooled about whether they are dealing with a machine, or whether something is real, you owe them a heads-up. That is the whole of Article 50, in one line.

These are transparency duties, not bans. Nothing here stops you using AI; it just asks you to be straight with people in four specific situations, and it has been in force since 2 August 2026. Most small teams are caught by one or two of them, and closing each is usually a small job, not a project. Want the whole map first, including which obligations apply to you at all and who enforces them in Luxembourg? Start with the full AI Act readiness guide.

The four things Article 50 actually covers

Find the row that sounds like you. If none of them do, you have nothing to label, and that is a perfectly good answer too.

If you... You must...
Run a chatbot or AI assistant that talks to people Tell users they are interacting with AI, from the first exchange, unless it is already obvious. Covers voice assistants and auto-reply tools too, not only chat windows.
Publish AI-generated or AI-edited images, audio, or video Disclose that it is AI-made where it is a realistic depiction that could be mistaken for authentic (a "deepfake"), and rely on machine-readable marking from the generating tool. Clearly artistic or obviously synthetic images are out of scope.
Publish AI-generated text on matters of public interest Disclose that the text is AI-generated. This is narrow: it targets unedited AI output presented as public-interest information, not ordinary marketing copy a human has reviewed and published.
Use emotion recognition or biometric categorisation Inform the people exposed to the system that it is operating. Rare for small organisations, but listed here for completeness.

And by when

The visible disclosures are live now: they have applied since 2 August 2026. There is no runway left on telling people they are talking to AI, or on disclosing realistic AI-made media.

One narrow exception on timing: generative systems already on the market before that date have until 2 December 2026 to add machine-readable marking to their output, and mainstream tools handle that part for you. How these dates sit against the delayed high-risk rules and the Digital Omnibus is covered on the AI Act readiness guide; this page stays on what you label and how.

What "labelling" looks like in practice

A real label is plain, and it sits where people actually look: not three clicks deep in a policy nobody opens. Three everyday examples:

  • A chatbot banner: a line at the top of the chat, such as "You are chatting with an AI assistant," shown before the conversation starts.
  • An image credit line: a visible caption near a realistic AI-made photo, such as "Image generated by AI," so a viewer is not misled about what they are looking at.
  • Content marking: the machine-readable metadata a generating tool embeds in its output. For mainstream tools this is done for you; your job is the human-visible disclosure on top.

The test is simple: would an ordinary person notice, and understand? Don't stick your disclosure in a place no regular person would scroll to. Just put it where they can see it.

What this does not require

Over-compliance has a cost too. Article 50 does not want you stamping "made with AI" on everything you touch like a nervous rubber stamp. You do not label the email you drafted with AI and then edited yourself. You do not flag a stylised illustration nobody would take for a photograph. And purely internal use, where no customer or member of the public is on the receiving end, sits outside these rules entirely.

The line is deception, not involvement. If a reasonable person could be misled about whether they are dealing with a machine, or whether something is real, say so.

Not sure which of these apply to you?

Run the 2-minute readiness check. It walks through how you actually use AI and tells you, in plain language, which obligations touch you and which you can ignore. It runs entirely in your browser, and nothing is stored.

Take the 2-minute readiness check

AI Act transparency: common questions

Do I have to label AI-generated content on my website?

Only some of it. Realistic AI-made or AI-edited images, audio, and video that could be taken for real must be disclosed; a chatbot must announce itself; AI-written text on matters of public interest must be flagged. Ordinary AI-assisted marketing copy and clearly artistic illustrations do not need a label.

Article 50 targets content that could mislead a reasonable person about what they are seeing or who they are dealing with. A photorealistic image of a real place or person that never happened is in scope. A stylised header illustration is not. AI help with a blog draft you then edit and publish under your own name is not the trigger either; the text rule is narrow and aimed at unedited AI output presented as public-interest information. When in doubt, check which of the situations above you are actually in, or run the readiness check.

Does my customer-facing chatbot need a disclosure?

Yes. Since 2 August 2026, anyone interacting with an AI system must be told they are dealing with AI and not a person, from the first exchange, unless it is already obvious. A one-line notice on the chat window is usually enough.

This is Article 50(1), and it is not limited to chatbots: a voice assistant or an automated reply tool that speaks to people on your behalf is covered too. If you built or configured the assistant yourself, the obligation is yours, not just your vendor's. In practice this is one of the quickest items to close: a visible line such as "You are chatting with an AI assistant" at the start of the conversation does the job.

What is the difference between the August and December 2026 deadlines?

The transparency duties themselves apply from 2 August 2026. The 2 December 2026 date is one narrow extension: generative AI systems already on the market before August have until then to add machine-readable marking to their output.

So there is no grace period on telling people they are talking to AI, or on disclosing a deepfake: those are live now. The December date covers only the technical marking obligation for pre-existing generative systems, and content produced before 2 August 2026 does not have to be marked retroactively. If you use a mainstream tool, its provider handles that marking; your job is the visible disclosure.

We only use AI internally: does Article 50 apply?

Generally no. The transparency rules are about not misleading the people on the other side of an interaction or a piece of content. Purely internal use, with no external audience being deceived, sits outside Article 50.

Drafting your own notes with AI, summarising internal documents, or running analysis your staff read does not trigger a labelling duty. The moment that output goes to customers or the public, whether a chatbot reply, a published deepfake-style image, or unedited AI text on a matter of public interest, one of the situations above applies. Note that a separate expectation, AI literacy for the staff using these tools, has applied since February 2025 regardless.

Sources

Every rule here comes from a primary EU or Luxembourg source:

For the wider legal framework, the Digital Omnibus, and Luxembourg supervision by the CNPD under Bill n°8476, see the sources on the AI Act readiness guide.

Regulatory content last reviewed 27 August 2026. This page is practical guidance, not legal advice.

Want the labels done properly, once?

We set the disclosures and content labelling up with you, so they are right and you can maintain them without calling anyone. On their own, or as part of wider AI Act readiness. And if you are not sure any of this even applies to you, that is a fine reason to talk too.

Book a call