Why a Prompt That Works in English Fails Silently in Another Language

You kept the instruction in English, added an output-language directive, and the Spanish output is fluent, on topic and correctly formatted. It is also measurably worse than the English output in ways nobody on the team can see, because every one of the failures below produces valid, confident text.

Why it is silent

Every failure in this catalogue has the same shape: the model resolves an ambiguity or an inapplicable constraint by doing something reasonable, and reasonable is not the same as what you asked for. There is no exception to raise, because nothing was violated — the model was given an under-specified instruction and it specified it.

That is why these are found by a native reviewer or by a targeted check and essentially never by monitoring. Latency is normal, the schema validates, the language detector is happy, the answer is on topic. If you are looking for a signal in your dashboards, the only one that usually moves is a slow drift in a downstream metric — resolution rate, thumbs-down rate, escalation rate — for one locale, months later.

Failures in the instruction

Idiom carried across the language boundary

An instruction like Flag anything that looks off. asks the model to interpret two English idioms and act in Spanish. It usually manages, but the interpretation is looser than it would be for a literal instruction, and the looseness shows up on the hard cases. The general rule: idioms in an instruction are a tax you pay only in the cross-lingual case, and the fix is free — write Mark anything inconsistent with the source document. and the instruction is now literal in every language.

Ambiguous reference the model resolves with English habits

Summarise the article and translate it. — translate the article or the summary? An English reader guesses from context, and so does the model, and its guess is shaped by the English text it saw most. The same ambiguity in a language that would drop the pronoun entirely gets resolved differently. Any instruction with an unbound it, them or the above is a candidate. Name the object.

A constraint that describes an English category

Do not use jargon. and Write at a sixth-grade reading level. both refer to calibrations that exist in English education and publishing. The readability formulas behind grade levels are defined on English syllable and sentence statistics and do not transfer; asking for a grade level in Finnish is asking for a number the model can only approximate by vibe. Say what you actually want — short sentences, common words, no acronyms — in terms that exist everywhere.

Failures in the format constraints

• Length in words. Under 100 words is not a portable unit. Chinese, Japanese and Thai do not delimit words with spaces, so the constraint has no surface definition; German compounding puts into one word what English spreads over four. Use characters, or a per-language token budget.
• Casing rules. Use title case for headings has no meaning in German, where nouns are capitalised regardless, and is actively wrong in French and Spanish, which use sentence case for titles. A model asked for title case in those languages will produce something that looks like an English translation.
• Punctuation the instruction assumes. Quotation marks differ — German uses low-high pairs, French uses guillemets with spacing, CJK uses its own brackets. An instruction that says put the term in quotes gets whichever the model favours, which is frequently the English ones.
• Numbers and dates. A decimal point where a comma belongs, a month-day date where day-month is expected, or Western digits where Arabic-Indic digits are conventional are all valid text and all wrong. These are covered separately in Arabic-Indic numerals in model output.
• Name order. An instruction to output first name, last name is a category error in languages and cultures where the family name comes first, and it silently produces reversed names for a subset of your users. See family-name-first name order.
• Negative constraints on a string. Never mention the word refund constrains an English string. The model writing Spanish will not write “refund”; it will write reembolso, and both your instruction and the filter you built to enforce it miss it entirely.

Failures in the examples

Few-shot examples are the highest-leverage part of a prompt and therefore the most damaging place to leave English assumptions.

A culturally specific example teaches more than its content. An example whose input contains a US street address, a phone number in (555) 123-4567 form, a date written 06/03/2026 and a sign-off of Dear Sir/Madam teaches the model that this is what the domain looks like. Given a German ticket, the model has been primed with a template that does not match, and it will sometimes reformat the user’s own data to fit — reversing a date it should have left alone, or restructuring an address into an order that no German postal system uses.

An English-language example next to a request for Spanish output has a subtler effect: the example demonstrates the mapping from input to output, and the model has to hold that mapping while changing language. The result tends to read like a translation of the English answer, because in a real sense that is what it is. The output is grammatical and slightly foreign, and this is the single most common cause of the complaint that the model’s Spanish “sounds translated”.

Both are fixed the same way: examples in the target language, with inputs that look like real inputs from that locale. That argument in full is in why few-shot examples belong in the target language.

Detecting them

These fail in ways monitoring cannot see, so the detection has to be deliberate. In rough order of cost:

1. Grep your own prompts for the constructions above. Idioms, unbound pronouns, word counts, title case, name order and English-string constraints are all findable by reading, and this pass usually finds several.
2. Add a per-language check for each format constraint you rely on: character length instead of word length, expected quotation marks, expected digit range, expected date pattern.
3. Translate your negative constraints into every supported language and check for all of them, not only the English one. The list is short and it is the highest-risk item here.
4. Build the English control into your evaluation so a per-language regression can be distinguished from a task failure. The harness is in testing whether a prompt works the same across languages.
5. Get a native speaker to read twenty real outputs per language, once per quarter. Nothing above substitutes for it, and twenty is enough to surface a systematic problem.

For languages with little representation in training data, all of the above still applies and a further set of problems arrives on top of it, including a higher rate of confident fabrication. Those are covered in prompting in a low-resource language.