How accurate is AI menu translation? What goes wrong with dish names

AI menu translation is usually fine for plain dishes and least reliable for the names that matter most: local, figurative and culture-specific ones. The few studies that test it find errors in both machine and human translations, with machine tools doing worse on culture-specific dish names. So use AI as a first draft, keep original dish names, translate plain descriptions, and have a native reader check every language you publish.

This post covers what the studies found, the six failures that come up most often, a 15-minute check you can run on your own menu, and where WaiterPilot’s chat fits (and where it doesn’t). If you are still choosing between a translated menu and a chat, start with translated menu or multilingual chat instead.

How accurate is AI menu translation?

Short answer: accurate enough for “grilled chicken with rice”, unreliable for dish names that carry local or figurative meaning.

Most of a menu is easy for a machine. “Chicken”, “rice”, “tomato sauce” and “grilled” have direct equivalents in most languages. The trouble sits in a small part of the menu: the dish names that only make sense inside a cuisine. That is also the part guests read first.

Researchers call these culture-specific items. Few studies test menu translation directly, and the ones we could open are small. Here is what each one found, with its scope:

Table: what three sources found about menu translation, as read on 16 Oct 2026

Two things stand out. First, human translators made plenty of mistakes too, so “pay a translator” is not a guarantee. Second, the company with the biggest menu-translation problem in the set still built a way to flag uncertain translations for people to look at. A small restaurant can do the same thing by hand in a few minutes.

What goes wrong with dish names?

Short answer: six things, and most of them happen to the name, not the description.

The Edinburgh paper opens with a good example. The Sichuan dish 蚂蚁上树 translates word for word as “Ants Climbing a Tree”. The authors say it “should be translated in English as ‘Sauteed Vermicelli with Minced Pork’”. The name is a picture: bits of pork on noodles look like ants on branches. A guest who reads “ants” will skip it, or order it for the wrong reason.

The six failures below are the ones the studies name, plus the ones a server sees at the table. The examples marked as ours are illustrations, not findings.

1. Word-for-word translation of a figurative name. “Ants Climbing a Tree”, “Buddha Jumps Over the Wall” (another example from the Edinburgh paper). The guest gets a riddle instead of a dish.
2. Underspecification. The translation is too vague to choose from. Our example: a Lithuanian or Polish dumpling dish turned into just “dumplings”, with no word on the filling. The KTU thesis lists underspecification among the most common AI errors.
3. Wrong cultural substitution. The tool swaps in a “nearest” foreign dish that isn’t the same thing. Our example: calling a local flatbread “pizza” sets the wrong expectation about size, topping and price. The thesis also lists inaccurate cultural substitution among the most common errors, for humans and AI alike.
4. A completely unrelated translation. Rare but real: the thesis counted 11 cases where AI output had nothing to do with the dish.
5. Dropped or changed ingredients. A description loses “with walnuts” or “fried in butter” on the way through. On a menu, this is the most dangerous error, because it is the one that touches allergies.
6. Names that don’t match across the restaurant. The translated menu says one thing, the ticket says another, and the server says a third. The guest ordered “lamb shoulder”, the kitchen sees “kleftiko”, and nobody is sure it’s the same plate.

The first four make the menu confusing. The fifth can make it unsafe. The sixth slows down every table that orders from the translation.

Why is the dish name the hard part?

Short answer: because a dish name is often a proper noun with a story, not a description, and translation tools are built for descriptions.

The Edinburgh authors sort culture-specific dish names into three levels:

The same three levels work for any cuisine. “Grilled sea bass” is concrete. “Patatas bravas” is creative: potatoes, plus a word (“brave”) that tells a Spanish speaker it’s spicy and tells everyone else nothing. “Pulpo a la gallega” sits between creative and abstract for a guest who doesn’t know Galicia.

That gives a simple rule for any translated menu: keep the original name, translate what is in it. A guest who reads “Patatas bravas: fried potatoes, spicy tomato sauce, garlic mayonnaise” knows what will arrive. They can also say the name to the server, who knows it.

How do you check a translated menu? A 15-minute checklist

Short answer: check the risky lines, not every line, and use one native reader per language.

You don’t need to proofread every word of a 60-dish menu. Most errors cluster in the names and in the ingredient lists. Spend your 15 minutes there, per language.

1. List your abstract and creative names. Go through the menu and mark every dish whose name isn’t a plain description. On many menus it’s only a handful.
2. Check each marked dish reads as “name + plain description”. If the translation replaced the name with a literal phrase, put the original name back.
3. Back-translate the descriptions of those dishes. Run the translated description back into your language with a different tool. If the main ingredient changes or disappears, rewrite it.
4. Read every allergen line against your kitchen’s list. Nuts, dairy, gluten, eggs, shellfish, sesame: the word must be there, in that language, on every dish that contains it.
5. Check cooking fats and hidden ingredients. Butter, lard, fish sauce, wine in sauces. These are the words that drop out.
6. Ask a native reader to scan the marked lines. A staff member, a regular, a student from a nearby language school. Ask one question: “Would you know what arrives?”
7. Check prices, sizes and units survived. Some tools reformat numbers or drop the currency.
8. Compare the translated name with the ticket name. Whatever the guest orders from, the kitchen should see a name it recognizes.
9. Date the check. Write the date and the reader’s name on the file. When the menu changes, the check is stale.
10. Repeat for each new dish. One new special means one new check, per language.

A worked example: Casa Olivo is a fictional restaurant (an example, not a customer). The figures are assumptions to show the time involved. Casa Olivo has 48 dishes and publishes translations in three languages. The owner marks 9 dishes as creative or abstract. Steps 1 to 5 take about 10 minutes per language for those 9 dishes, and the native-reader scan (step 6) takes 5. That is 15 minutes per language, or 45 minutes for three, compared with proofreading all 48 dishes three times over. Every new special adds a minute or two per language.

Is a human translator more accurate than AI?

Short answer: in the one direct comparison we found, yes, but not by enough to skip checking.

In the KTU thesis, human translators made 101 errors in 154 translations of the same 30 Lithuanian items; the AI apps made 135 in 118. Errors are counted per problem, so one translation can hold more than one. The thesis concluded that human translation was more reliable for those items.

The practical reading for an owner: a paid translation is a better first draft, and it still needs the same check for names and allergens. Neither replaces a reader who knows the cuisine.

Where does WaiterPilot fit?

Short answer: WaiterPilot doesn’t translate your menu as a document; its chat answers guests’ questions about your menu in their language.

Be exact about what changes language and what doesn’t:

Table: language of each part of WaiterPilot

This avoids some of the six failures and not others.

• Figurative names (1) and mismatched names (6): dish names stay as written on your menu, on the guest’s screen and on the ticket, so the kitchen sees the name it knows. In our 15 Oct 2026 test, a Spanish question was answered in Spanish with dish names kept as written on the menu. Tapping a dish name in the chat opens its photo, which answers “what is it?” without words.
• Underspecification (2) and substitution (3): the chat explains a dish from your own description. If your description is thin, the answer will be too. When the menu doesn’t say, the chat tells the guest to ask a member of staff rather than guess.
• What it doesn’t remove: nobody proofreads each reply before the guest reads it. That is why we name only the languages we have tested, and why you should try the languages your guests use before relying on them. The multilingual ordering page shows the same question answered in English, Spanish and Gulf Arabic.
• Notes: a special request is written by the AI from the guest’s words and is not translated for staff. A note can reach the kitchen in the guest’s language. Decide who reads those, usually the server who confirms the table’s first order. This matters in many kitchens: 30% of US restaurant and foodservice employees speak a language other than English at home (NRA, 2026, from Census data).
• Arabic: the owner’s menu editor shows how much of the menu an Arabic-speaking guest reads in Arabic (sections, dish names, descriptions, choices). Anything without Arabic text shows in English. Our guide to writing a menu in Arabic covers how to fill it.

On allergies, WaiterPilot gives what your menu says and asks the guest to confirm with staff. Your staff always have the final word. See allergies and dietary questions.

When is a translation tool the better choice?

Short answer: when guests want to browse the whole menu in their language without asking anything.

If most of your foreign guests want to read quietly and point, a translated menu serves them better than any chat. As published on 15 Oct 2026, MENU TIGER auto-translates menus “into 30+ foreign languages” and has a free plan (MENU TIGER pricing), which is less than WaiterPilot’s $200 per location per month. Run the checklist above on whatever it produces.

WaiterPilot makes more sense when the problem is the follow-up questions (“is it spicy?”, “what comes with it?”) and your servers spend the rush answering them. Our comparison of the options lays out both, and serving tourists who don’t speak your language has a full checklist.

What to do next

1. Mark your creative and abstract dish names today.
2. Rewrite each as “original name: plain ingredients”.
3. Run the 15-minute check on the language your guests use most.
4. Put allergen words in every language you publish, and keep the rule that staff decide.

To see how a chat handles questions about your dishes in the guest’s language, try a sample order. WaiterPilot costs $200 per location per month after 30 days free, with 0% commission; the details are on the pricing page.