1. The deadline that is still open
The transparency obligations of Article 50 have applied since 2 August 2026. For systems placed on the market before that date a transitional period runs — and it ends on 2 December 2026.
In practice that affects almost every property that has been working with a chatbot, a text tool or an image generator for some time. Anyone who concluded in the summer that their own solution was, after all, older was not wrong — but the grace period is running out. Three months are no reason to rush and no reason to let it lie.
The general rules — the four categories of case, the range of fines, supervision — are in our article EU AI Act Article 50: AI disclosure in marketing. Here it is about application in the property.
2. You are almost always a deployer, not a provider
This distinction decides half of your obligations, and it is regularly made wrongly.
A provider is whoever develops an AI system or places it on the market in the EU. Those are your tool suppliers — the manufacturer of the chatbot, the vendor of the image generator, the operator of the language model.
A deployer is whoever uses such a system under their own responsibility. That is you.
The difference in practice: the machine-readable marking of synthetic content is owed by the provider. The visible disclosure to your guests is owed by you. Both together make up the chain — and you cannot rely on the first part taking care of the second.
3. The booking chatbot
The most common case and the simplest. If an AI system answers on your website, on WhatsApp or in a messenger, the person must know — unless it is obvious from the circumstances.
"Obvious" is a narrow term. A window that opens with "Hello, I'm Marie from reception" is the opposite of obvious. A note in the first sentence or right at the input field is enough and costs nothing:
- At first contact, not in the legal notice. The disclosure belongs where the conversation begins.
- Understandable, not legalistic. "You are chatting with an AI assistant. For personal queries I will pass you on to our reception" is sufficient.
- A way to a human. Not an obligation under Article 50, but the obvious addition — and in hospitality a matter of course anyway.
The same applies to voice assistants on the telephone. A property that lets calls be answered by a voice that sounds human has a disclosure case, not a service advantage.
4. Images: mood yes, rooms no
This is where marketing parts company with misleading, and the AI Act is only one half of the problem.
Legally the labelling obligation applies to AI-generated or manipulated images. Providers must mark machine-readably; where an image looks deceptively real and shows people or events, the visible label is added for you as deployer.
Practically the more important boundary is a different one: a generated image showing a room, a view or a dish that does not exist in that form is a problem regardless of the AI Act — warranty, competition law, and above all the review the guest writes afterwards.
The line that proves itself in everyday practice:
- Real photographs for everything the guest must be able to recognise: rooms, facilities, views, food.
- Generated images at most for abstract mood motifs, backgrounds, illustrations — and then labelled.
- Retouching remains retouching. For standard editing the AI Act expressly provides an exception; but anyone who retouches a building away or adds a beach leaves that area.
5. Texts: editorial control lifts the obligation
The point that yields the most for your content work — and is the least known.
For texts on matters of public interest — health, environment, politics and the like — deployers face a labelling obligation if no editorial control has taken place. If it does take place and a human assumes responsibility for the content, the obligation lapses.
For hotels this means two things:
- The usual house content mostly does not fall under it anyway. Room descriptions, directions, menu texts and offer pages are not matters of public interest.
- Where it does apply — a guide to allergens, a piece on sustainability in the business, a statement on a regional issue — editorial control is the clean route. Not as a formality but as what it is: someone reads the text, checks the statements and stands behind it.
That is no burden but a description of good editorial work. Anyone who proofreads anyway has dealt with the case — and should record that it happened and by whom.
6. Replies to reviews: the underestimated case
Replies to reviews are increasingly generated, and they are the case where the most can go wrong.
From the point of view of Article 50 they are usually not texts of public interest, and with editorial control the question does not arise anyway. The risk lies elsewhere: a generated reply that does not engage with a genuine complaint does more damage than no reply at all. And the platforms have their own rules on automated content, which the AI Act leaves untouched.
The workable rule: generate as a draft, approve as a human, and for every complaint that names a specific incident, write by hand.
7. What labelling requires technically
The AI Act requires markings that are effective, reliable, robust and interoperable — and machine-readable. For deepfakes and texts subject to labelling, the label must additionally be perceptible without special technical aids, that is visible or audible.
For your property this means, in descending order of importance:
- Visible notices where people need them: at the chatbot, under a generated image, at the start of a labelled text.
- Machine-readable marking — owed by the provider. Your job is not to destroy it: anyone who pushes generated images through converters or has metadata stripped on upload deletes the marking that carries the proof.
- Record keeping. Who uses which tool, who proofreads, when it was checked. One page is enough.
8. The Code of 10 June: voluntary, but useful
On 10 June 2026 the Commission published the Code of Practice on the transparency of AI-generated content. By the end of July around 190 companies and organisations had signed.
It is voluntary — the obligations behind it are not. The practical value: signatories can cite the measures of the Code as evidence of their compliance and have an easier time with supervisory authorities. Non-signatories have to have their own measures assessed individually.
For an individual property signing is rarely the obvious step. For chains, booking platforms and tool providers it is worth a look.
9. What to do by 2 December
Seven steps, in this order, achievable in one morning for a mid-sized property:
- Take stock. Where in the property does AI work? Chatbot, telephone, texts, images, replies to reviews, translations, price recommendations. A list, not an analysis.
- Per entry: ask the provider. Does the tool mark machine-readably? Answer in writing, to the records.
- Disclose chatbot and telephone voice. One sentence at first contact. That is the quickest item on the list.
- Sort images. What is generated, what is real? Take generated material out of everything the guest must be able to recognise.
- Lay down editorial control. Who proofreads before something is published — and where is that noted.
- Protect metadata. Check image editing and upload routes to make sure they do not remove markings.
- Write it down once. One page: tools, responsibilities, checking steps, date. That is your record.
Legal note, not legal advice: this article summarises publicly accessible sources and does not replace a review by a lawyer. For deepfakes, the presentation of reviews and price information, competition, trademark and contract law additionally apply.
10. FAQ: frequently asked questions about the AI Act in hotels and hospitality
From when does the labelling obligation apply to my hotel?
The transparency obligations of Article 50 have applied since 2 August 2026. For AI systems placed on the market before that date a transitional period runs until 2 December 2026.
Do I have to label my chatbot as AI?
Yes, unless it is obvious anyway. The notice belongs at first contact, not in the legal notice, and must be understandable to an ordinary guest. The same applies to voice assistants on the telephone.
May I use AI-generated images in hotel brochures?
Legally the labelling obligation applies; practically the tighter boundary is a different one. Everything the guest must be able to recognise — rooms, facilities, views, food — should be shown as a real photograph. Generated motifs come into question at most for abstract mood images, and then labelled.
Do I have to label AI-written blog texts?
Only for matters of public interest such as health or the environment — and even there only if no editorial control has taken place. If a human proofreads and assumes responsibility, the obligation lapses. Room descriptions and offer pages do not fall under it anyway.
How high are the fines?
The range for breaches of the transparency obligations goes up to 15 million euros or 3 per cent of worldwide annual turnover. The size of the company is taken into account.
Should we sign the Code of Practice?
For an individual property usually not necessary — the obligations apply anyway, and the Code is aimed above all at providers and larger deployers. For chains, booking platforms and tool providers signing can simplify the evidence.
11. Sources and status
Status: September 2026. The basis is the publications of the European Commission.
- Transparency obligations under Article 50 of the AI Act — questions and answers from the European Commission — source for the start of application, the transitional period until 2 December 2026, the four categories of case, the exceptions, the technical requirements and the range of fines.
- Code of Practice on Transparency of AI-generated Content — source for the publication date of 10 June 2026, the voluntary nature and the number of signatories.
Fundamentals on the topic: EU AI Act Article 50 — AI disclosure in marketing. On data protection when using tools: Using AI tools under the GDPR: what you need in writing.
