AI on your menu: what the new EU rules actually require

By Cravd Studio · 7 September 2026 · 9 min read

AI is already on most menus — in the photography, in the translations, sometimes in the dish descriptions themselves. On 2 August 2026 the EU's transparency rules for AI-generated content began to apply, with a second deadline on 2 December. If you run a restaurant, the direct obligations are narrower than the headlines suggest — but the practices worth adopting are broader. Here is the honest version.

What changed on 2 August 2026

Article 50 of the EU AI Act — the transparency chapter — started applying on 2 August 2026. Two duties matter here. Providers of generative AI systems must mark synthetic output so it can be detected by machine. Deployers must clearly label deepfakes, and AI-generated text published to inform the public on matters of public interest. Systems already on the market before that date have until 2 December 2026 to implement the machine-readable marking. The Commission has also published a voluntary Code of Practice on the transparency of AI-generated content as a recognised route to compliance. Penalties for non-compliance run to €15 million or 3% of worldwide annual turnover.

Does any of it apply to a restaurant menu?

Mostly not directly, and it is worth being precise rather than alarmed. A stylised photograph or short video of your own dish is not a deepfake: that term concerns content resembling real people, places or events in a way that would mislead. A menu is also not "text published to inform the public on matters of public interest" — that provision is aimed at news and civic information, not at describing a burrata. And the machine-readable marking duty falls on the provider of the AI system you use, not on the restaurant buying the output.

One case does land squarely on you: if you put an AI assistant in front of guests — a chatbot taking bookings, an automated phone or text ordering line — people must be told they are dealing with an AI. That is an explicit Article 50 duty for the deployer, and it is increasingly common in hospitality.

The rules that were already binding

The AI Act is the new arrival, but it is not the rule most likely to catch a restaurant out. Two older ones bite harder. Under EU Regulation 1169/2011, food information must not mislead as to the nature, identity, qualities or composition of the food — and pictures are food information. Consumer protection law says much the same in broader terms: you may not mislead a customer into a transaction they would not otherwise have made. In the UK, the CAP Code and the ASA apply the same standard to advertising.

So the legal test is not really "was AI involved?" It is "does this honestly represent the dish you will serve?" A heavily retouched photograph taken by a human can fail that test. A carefully produced AI-assisted video of the real plate can pass it. The technology is not the point; the accuracy is.

Where AI actually shows up on menus today

Four places, in rough order of how common they are. Imagery: enhancing, relighting or animating dish photographs. Translation: rendering the menu into other languages. Descriptions: drafting or tightening the words for each dish. Assistants: chat and voice ordering. Each carries a different risk, and only the last has an unambiguous labelling duty attached to it today.

Translation is the one to be careful with

If you take one operational point from this article, take this one. Machine-translating allergen information written as free text is genuinely dangerous. "Nut-free" can come back from a translation engine as something a guest reads as "nuts" — and on a safety-critical field there is no acceptable error rate. The fix is structural, not linguistic: store allergens as a fixed set of numbered codes, identical for every dish, and render the wording for each language from a legend a human has checked. Translate the prose around the dish freely; never translate the safety data as prose.

The same logic applies to anything a guest could act on. Descriptions and marketing copy can be machine-translated with a light human review. Allergens, dietary flags and anything a guest with a medical condition relies on should be structured data with a person in the loop.

What good practice looks like now

Six things, none of them onerous. One: label AI-assisted imagery or motion where a guest could reasonably be misled about what they are looking at — a short, unfussy note is enough. Two: never let a generated image invent the dish; it should be produced from a real photograph of the plate you actually serve. Three: keep a human approval step before anything reaches a guest. Four: keep allergens as structured, human-confirmed codes, never machine-translated free text. Five: if you deploy an AI assistant, disclose it. Six: ask your suppliers whether their systems will carry machine-readable marking by 2 December 2026 — that is their obligation, but it is your menu.

How we handle it at Cravd

We are in this category ourselves, so it is fair to say what we do. Every dish video starts from one professional photograph of the real plate, supplied by the restaurant — the still is used as-is, and the motion is generated from it. Nothing is invented: if the dish is not photographed, there is no video. The restaurant reviews and approves each dish before it goes live, and the guest app labels the result as AI-generated motion rather than passing it off as filmed footage. Allergens are numeric codes confirmed by a person before publication, never free text, and never machine-translated. That combination is what lets a menu be translated into eight languages without the safety information drifting.

The question a guest cares about is not whether AI touched your menu. It is whether the menu tells the truth about the food.