Do you have to label AI content? For a restaurant, mostly no.

Does this even apply to you?

The AI Act is EU law, but "I'm not in the EU" is not the end of the sentence. Article 2 extends it to businesses established outside the Union where the output produced by the AI system is used in the Union. That is a wider net than most people expect.

Three sentences and you know where you stand.

One article. Not the whole Act.

The EU AI Act is a large piece of law and most of it has nothing to do with you. The bulk of it governs high-risk systems — AI in medical devices, in large-scale recruitment, in law enforcement. For a restaurant, exactly one article matters: Article 50, the transparency obligations. It has applied since 2 August 2026.

Worth knowing, because it is being reported badly: the Digital Omnibus that entered into force at the end of July 2026 did push several AI Act deadlines back. It moved the high-risk obligations to late 2027 and 2028. It did not touch the deployer duties in Article 50. If you have read that AI labelling was delayed, that is wrong. (One technical deadline did move, and it belongs to the tool vendors: providers who put a generative system on the market before 2 August 2026 have until 2 December 2026 for the machine-readable marking. Nothing changes for you.)

Provider or deployer?

The Act defines two roles, and nearly every misunderstanding comes from mixing them up.

A provider builds the AI system and puts it on the market: OpenAI, Google, Canva, Midjourney. Providers have to mark their outputs in a machine-readable format so software can later detect them as AI-generated. That is Article 50(2), and it is not your job. You do not need to embed anything in a file, add watermarks, or manage metadata.

A deployer is you: someone using an AI system in the course of business. Deployers are covered by Article 50(4), which is short. It contains exactly two duties — one for image, audio and video, one for text.

Your restaurant, gone through properly.

This is the part worth printing out. Every AI use that actually comes up in a restaurant, with a clear verdict.

Text

The reason so much of this is green is the wording itself. The text duty covers text published with the purpose of informing the public on matters of public interest. The Commission's FAQ spells out what that means: politics and democratic processes, public administration, justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and any economic, financial, political, scientific or cultural developments that could be a relevant subject of public debate. Your daily special is none of them. Note that this is a list of examples, not a closed one — anyone publishing regularly on nutrition or sustainability should look twice.

And even if a text did fall inside, the exemption follows immediately: no duty where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication. The Commission is explicit that a spellcheck does not count — it has to be a real look at the substance. Which is what you already do before hitting publish.

Images, video, audio

The Act calls this a deep fake, and the word misleads — it sounds like political videos, but the definition is sober: content that resembles existing persons, objects, places, entities or events and would falsely appear authentic. The Commission’s guidelines break that into three tests that must all hold — close resemblance, something that exists or could plausibly exist, and the false impression of authenticity. A photorealistic burger that never came out of your kitchen ticks all three. Photorealism alone isn’t decisive, though: the Commission looks at the whole picture — resemblance, message, context and what the audience expects.

The point most people miss: there is no "but I checked it myself" exemption for images. The editorial-review carve-out sits only in the paragraph about text. With images, your only protection is that they are visibly artificial — or a label.

The paragraph does soften the duty for evidently artistic, creative, satirical or fictional works. Don’t lean on that as a business: the Commission reads it narrowly for advertising, and where informative and creative elements mix, the informative character prevails. An ad photo stays an ad photo.

When AI talks to your guest

Strictly, the chatbot duty falls on the provider of the system first. That helps you very little in practice: your logo is on the page and the guest comes to you. The exemption only bites where it is obvious anyway, judged by a reasonably well-informed, observant person. A chat window with a human name and a profile photo is the opposite of obvious.

Everything that stays internal

The whole internal side is clear. Which is also where AI actually saves a restaurant money — and precisely there, no labelling question arises at all.

How to actually write it.

The Act prescribes no fixed wording. It requires the disclosure to be clear and distinguishable, no later than the first interaction or exposure, and perceivable without any special tools. A metadata entry expressly does not count — a human has to be able to see or hear it.

Images

On the image itself, not three paragraphs below it, and not buried in a caption that Instagram collapses. What works:

Created with AI
AI-generated image
Illustrative image, AI-generated
Image: AI-generated

On social, put it visibly in the image or at the very start of the caption. Instagram and Facebook also have their own AI flag you can set on upload — useful, but it does not reliably replace your own disclosure, because it renders differently depending on where the post is seen.

Chatbot

First line in the window, before the guest types:

Hi! I'm the digital assistant at [name] — an AI, not a person.
I can help with opening hours, bookings and questions about the menu.
If you'd rather speak to someone from the team: [phone number].

That last line isn't legally required, but it saves you guests. People who realise they're talking to a machine usually want to know immediately how to get out of it.

Phone

Hello, this is the automated booking assistant at [name].
I'm a voice system. What date would you like to book?

Deliberately at the start, deliberately in one sentence. Push it to the end and you have technically complied while annoying the guest anyway.

What can actually happen.

This is where the internet exaggerates most, so here are the numbers straight. Article 99(4) sets the ceiling for Article 50 breaches at 15 million euro or 3 percent of total worldwide annual turnover, whichever is higher. That is the figure in the headlines.

Paragraph 6 of the same article reverses it for small and medium enterprises: there, the lower of the two applies. For a business turning over two million euro, the ceiling is 60,000 euro. Still real money — but a different order of magnitude from what compliance vendors are currently advertising, and it is a worst-case maximum, not a standard penalty.

One caveat almost every article omits: that only holds if you actually are an SME. The EU definition applies — under 250 staff and turnover no more than 50 million euro — and linked enterprises count towards it. A restaurant owned by a larger group or chain can fail that test, and then the higher figure applies again: 15 million. It’s a cliff edge, not a sliding scale.

Enforcement is national and uneven. Germany named the Bundesnetzagentur at the end of July 2026 and it runs a free AI service desk aimed explicitly at small businesses. Spain has AESIA, though its national AI law was still a bill in Congress as of August 2026. France and Austria had not formally designated an authority at all. That does not mean the obligations aren't in force — they are, everywhere — only that who knocks on the door varies.

UK, US and everywhere else.

United Kingdom. No AI Act, and no general legal requirement to disclose AI in advertising — the ASA's position is that the existing rules apply regardless of how content was made. What bites instead is misleadingness. An AI image of a dish that doesn't reflect what you serve is a CAP Code problem exactly as a heavily retouched photo would be, and the ASA has been blunt that labelling something as AI does not cure a fundamentally misleading message.

The sharper regime is the Digital Markets, Competition and Consumers Act, in force since April 2025. Fake reviews are explicitly banned — writing them, commissioning them, hiding incentivised ones — and AI-generated reviews are fake reviews. The CMA can now decide infringements itself, with penalties reaching 10 percent of global turnover. That is a far bigger number than anything in the AI Act, and it is aimed squarely at exactly the thing some restaurants are tempted to use AI for.

Switzerland. Not an EU member, no AI act in force. The Federal Council decided in 2025 to ratify the Council of Europe AI convention and adjust sector rules rather than write a Swiss AI Act; a consultation draft is expected around the end of 2026. What applies today is unfair-competition law — the ban on misleading statements, which needs no AI-specific rule — plus guidance from the data protection commissioner that manipulating an identifiable person's face or voice must always be clearly recognisable.

United States and elsewhere. No federal labelling duty. State law varies and moves fast. The constant everywhere is that consumer protection authorities treat a misleading image as a misleading image, whatever made it — and that fake reviews are enforced hard.

"Anything made with AI has to be labelled."

That sentence is wrong, and it does damage. It makes operators either stop touching AI out of fear, or stamp "AI-generated" under every piece of copy as a precaution — which makes them look worse than they need to.

The Act does not ask was AI used here? It asks is someone being misled about something they ought to know? That is why the drawn image is free and the photorealistic one isn't. Why the chatbot is covered and the invoice analysis isn't. And why the ad copy you read over yourself needs no label — you're putting your name to it, exactly as you did when an agency wrote it for you.

The honest summary: if you use AI to work faster and a human looks at it before it goes out, you're fine. If you use AI to show something that doesn't exist, say so.

Questions, answered.