Prompting AI in Languages Other Than English

TL;DR: Prompt language and output language are separate choices, not one. Many models reason more reliably in English even when you need the final answer in Spanish, German, Arabic, or Japanese. State the output language, register, and any do-not-translate terms explicitly, since models drift back to English without that instruction, and quality still varies sharply by language.

Should You Prompt AI in Languages Other Than English?

The short answer: it depends on the task, but the language you type in and the language you want back are two settings, not one automatic pair.

If you write a prompt in Spanish, most current models answer in Spanish by default, matching your input. That's fine for everyday requests like a business email, a social post, or a product description. But for reasoning-heavy tasks (analysis, code, math, structured extraction), research keeps finding the same shape of result: models still process instructions more reliably in English internally, even when the final answer needs to land in another language.

OpenAI measured this directly in the GPT-4 technical report, on the MMLU benchmark: English scored 85.5% accuracy, against German 83.7%, Latvian 80.9%, Swahili 78.5%, and Welsh 77.5% (OpenAI, GPT-4 Technical Report, arXiv:2303.08774, March 2023). That's three-year-old data on a model two generations behind current ones, and the gap has likely narrowed since. No lab publishes a directly comparable, recently dated breakdown on a predictable schedule, so treat the direction as more durable than the exact numbers.

More recent work on lower-resource languages found the same shape held: a 2024 evaluation of GPT-4, Llama 2, and Gemini across Bangla, Hindi, and Urdu found all three scored better prompted in English than in the local language, across every format tested (Better to Ask in English, arXiv:2410.13153, October 2024).

None of this means always prompt in English. For reasoning-heavy work in a lower-resource language, write the prompt in English and explicitly request output in your target language. For everyday tasks in a well-supported language, prompting entirely in that language is fine and skips a step you don't need.

Our ChatGPT prompt framework guide covers the Role/Task/Format/Constraints/Tone structure that every instruction below slots into.

How Do You Tell an AI Model Which Language to Reply In?

State it as its own line, right before the actual task, not folded into a sentence in the middle of a longer prompt.

Output language: German (formal register — Sie, not du)

Task: Draft a reply to a customer whose shipment is delayed by
two weeks. Explain the reason, apologize once, and offer a
10% discount on their next order.

Models generally match your input language by default for short, simple requests. That default gets less reliable as a prompt gets longer, references pasted-in material, or involves multi-step reasoning, which is exactly when a mismatch is most costly, since you're least likely to catch it without reading the target language fluently.

Research on long chain-of-thought reasoning found a mechanism behind this drift: as reasoning gets longer and more complex, models increasingly shift back toward their dominant pretraining language, usually English, even when explicitly asked to reason in the target language, a pattern the authors call "cross-lingual collapse" (Cross-lingual Collapse, arXiv:2506.05850, submitted June 2025, updated February 2026).

Whether this instruction lives in the system prompt or the user prompt matters less than that it exists and sits close to the actual request. In a reusable template, put the language line at the end, right before the variable content. That's the position models weight most heavily.

This is also where a reusable setting earns its keep. Prompt Architects' Global Variables and Personal Context Library exist for exactly this kind of instruction you'd otherwise retype every time. Store your output language and register once, and every prompt you enhance can carry it automatically.

How Do You Get the Right Register — Tú vs Usted, Du vs Sie, Keigo?

Name the exact register. Don't leave it to the model's default.

• Spanish: tú (informal) vs usted (formal), plus Spain's vosotros vs Latin America's ustedes for the plural "you."
• German: du (informal) vs Sie (formal).
• French: tu (informal) vs vous (formal).
• Japanese: casual speech vs keigo (敬語), itself split into teineigo (polite), sonkeigo (respectful/honorific), and kenjōgo (humble).

We looked for a dated, rigorous benchmark measuring which register models default to across languages, and didn't find one solid enough to cite a number. Practitioner reports describe a moderately formal default, but that's observed, not measured. Treat any claim about the "default," including this one, as unverified, and specify what you want.

Actúa como asesor financiero senior. Redacta un correo dirigido
a un cliente explicando el retraso de su pago de intereses este
trimestre.

Registro: formal (usted, no tú). Español de España, no
latinoamericano.

Du bist Fitness-Coach für Einsteiger. Schreib eine kurze,
motivierende Nachricht an einen Kunden, der drei Trainings in
Folge verpasst hat.

Ton: locker, freundlich, Du-Form. Antworte auf Deutsch.

あなたはカスタマーサポート担当者です。故障した製品について
謝罪し、返金対応を案内するメールを書いてください。

敬語（丁寧語・謙譲語）を使用し、失礼のない言い回しにしてください。

The Spain-vs-Latin-America and formal-vs-informal splits are the two mistakes we see most: a model asked simply for "Spanish" or "French" will guess, and won't always guess the audience you meant.

How Do You Stop AI From Translating Your Brand Name?

Add an explicit do-not-translate list. Without one, models sometimes localize or transliterate proper nouns (brand names, product names, technical terms), especially into a language with a different script, like Arabic, Japanese, or Hebrew.

You are a technical writer. Rewrite the following release notes
for a Spanish-speaking audience.

Output language: Spanish (Latin American, tú form)

Do not translate these terms — keep them exactly as written:
- Prompt Architects
- Chrome extension
- API
- webhook
- JSON

Release notes: [paste text here]

If your brand and technical vocabulary don't change often, save that exact list as a Global Variable or Personal Context entry so every future prompt inherits it instead of you rebuilding it by hand.

What Should You Know About Prompting in Arabic or Hebrew?

Two different problems get conflated here. The first is model output quality: the same question as any other language, depending on the model and how much Arabic or Hebrew training text it saw. The second is interface rendering, a UI problem, not a model problem, and a real, documented one. Multiple 2026 write-ups, plus an open thread on OpenAI's own developer forum, describe ChatGPT's input box failing to apply right-to-left alignment to Arabic and Persian text, with mixed RTL/LTR content — an Arabic sentence quoting an English function name, say — rendering with the Latin fragment in the wrong position (ChatGPT Arabic Guide: RTL Support & Tips, ai-toolbox.co, 2026; OpenAI Developer Community, "Issue with Right-to-Left Text Alignment in the Prompt Box", thread open as of the date cited). Several browser extensions exist specifically to patch this.

For do-not-translate terms in an RTL sentence, say explicitly that the Latin-script term stays as-is, because models sometimes transliterate a brand name into Arabic or Hebrew characters unless told not to:

أنت مستشار تسويق. اكتب منشورًا قصيرًا للترويج لمنتج جديد.

حافظ على اسم العلامة التجارية "Prompt Architects" مكتوبًا
بالأحرف اللاتينية كما هو، دون ترجمة أو كتابته بالعربية.

אתה כותב תוכן שיווקי. כתוב פסקה קצרה על יתרונות כלי ניהול
זמן חדש.

שמור על שם המוצר באנגלית, ללא תרגום או תעתיק לעברית.

Does Mixing Languages in One Prompt Help or Hurt?

It depends on whether the mixing is deliberate. Unplanned code-switching (languages mixing mid-sentence in a way the model has to parse on the fly) measurably hurts. A 2025 evaluation of code-switched text found LLMs show systematic degradation in grammaticality, factuality, and safety behavior when foreign-language tokens are inserted into an English matrix sentence, regardless of which language got mixed in (Lost in the Mix: Evaluating LLM Understanding of Code-Switched Text, arXiv:2506.14012, 2025).

Structured, deliberate mixing is different. Research on code-mixing found translating a prompt to English first, or mixing languages inside labeled few-shot examples, can act as a bridge that improves performance on lower-resource languages; it helps precisely because it's controlled and legible, not accidental (Code-Mixer Ya Nahi, arXiv:2410.11079, 2024).

That gives two legitimate reasons to mix languages on purpose: translate-first, for a lower-resource target language (write instructions in English, request output in the target language; that's sequencing, not really mixing), and bilingual audience content, where readers code-switch themselves and you name the intended mix explicitly, like a Spanglish post for a bilingual U.S. audience, rather than letting the model improvise it.

What doesn't work is a prompt that randomly switches languages mid-thought and hoping the model sorts it out. That's the exact pattern "Lost in the Mix" measured degrading.

14 Copy-Paste Prompts for Prompting in Other Languages

These cover every pattern above. Swap the bracketed placeholders for your own task.

1. Reusable template: the pattern to copy for any new language

Prompt language: English
Output language: [insert — e.g., Portuguese, Korean, Swahili]
Register: [formal/informal — name the local equivalent, e.g.
"usted" not "tú", or "Sie" not "du"]
Do not translate: [brand names, product names, technical terms]

Task: [your actual request]

2. Explicit output-language instruction, minimal version

Output language: French. Everything else stays as written.

Task: Summarize the attached quarterly report in three bullet
points for a non-technical board member.

3. Formal Spanish (usted), Spain

Actúa como un asesor legal. Redacta una carta breve solicitando
una prórroga de plazo a un cliente.

Registro: formal (usted). Español de España.

4. Formal Spanish, Latin America, contrasted with Spain

Eres redactor publicitario. Escribe un anuncio de 40 palabras
para una app de productividad, dirigido a un público de México
y Colombia. Usa "ustedes", no "vosotros", y evita modismos de
España.

5. Informal German (du)

Du bist Social-Media-Manager für eine Fitness-App. Schreib
einen kurzen, lockeren Instagram-Post zur Ankündigung eines
neuen Features. Du-Form, freundlicher Ton.

6. Formal German (Sie), business context

Sie sind Assistent eines Rechtsanwalts. Verfassen Sie ein
Antwortschreiben an einen Mandanten zur Fristverlängerung.

Sie-Form, sachlicher und formeller Ton.

7. Japanese with keigo specified

あなたはカスタマーサポート担当者です。配送遅延について謝罪する
メールを書いてください。

丁寧語を使用し、丁寧で誠実な印象を与える文章にしてください。

8. French, English-pivot reasoning

Réfléchis étape par étape en anglais, puis rédige ta réponse
finale entièrement en français.

Tâche : explique la différence entre un CDI et un CDD à un
nouvel employé.

9. Do-not-translate list, applied to documentation

Rewrite this help-center article for a German-speaking audience.

Output language: German (Sie form)
Do not translate: Prompt Architects, Chrome extension, API,
webhook, JSON, MCP server

Article: [paste text]

10. Arabic, brand name preserved in Latin script

أنت مستشار تسويق. اكتب منشورًا قصيرًا على وسائل التواصل
الاجتماعي للترويج لمنتج جديد.

حافظ على اسم العلامة التجارية "Prompt Architects" باللاتينية،
دون ترجمة.

11. Hebrew, product name preserved

אתה כותב תוכן שיווקי. כתוב פסקה קצרה המסבירה את היתרונות של
כלי ניהול זמן חדש.

שמור על שם המוצר באנגלית, ללא תרגום או תעתיק.

12. Deliberate bilingual (code-switched) content, named explicitly

Write a short Instagram caption for a bilingual U.S. Latino
audience that naturally code-switches between English and
Spanish, the way a bilingual speaker actually would — not a
literal translation into either language alone.

13. Prompt entirely in the target language (no pivot needed)

Eres un experto en marketing digital. Escribe tres ideas de
campañas de correo electrónico para relanzar un producto de
suscripción cuyo uso bajó un 20% este trimestre. Responde
completamente en español, tono profesional pero cercano.

14. JSON output with non-English string values

Output the response as JSON with this schema:
{ "subject_line": string, "body": string, "cta": string }

All string values in Portuguese (Brazil, informal register).
Keep the JSON keys in English exactly as shown.

Task: Draft a re-engagement email for a customer who hasn't
logged in for 30 days.

For more on the JSON prompt structure itself, see JSON Prompts Explained. The same output-language and do-not-translate rules apply to string values inside a schema.

Where Does Multilingual Prompting Still Fall Short?

Quality still varies sharply by language, and English-as-pivot is not a universal fix.

Low-resource languages score materially worse across every published multilingual benchmark we found, simply because there's less training text available in them. That gap has narrowed over the past several years but hasn't closed, and it won't close evenly. German or Japanese sits in a different position than Swahili, Latvian, or Welsh, and no single tip here erases that.

The "prompt in English, reason in English" trick isn't universal either. Research on long chain-of-thought reasoning across nine languages found English-as-pivot reasoning gave no benefit for French, improved performance for Japanese and Latvian, and proved insufficient for Swahili; the benefit is language-specific (Long Chain-of-Thought Reasoning Across Languages, arXiv:2508.14828, submitted August 2025, accepted to ICLR 2026). Don't assume the pivot helps your target language without testing it both ways.

Models change fast enough that any number in this post, including the ones cited above, is a snapshot, not a permanent ranking. If output quality is load-bearing (legal, medical, financial, anything where an error matters), have a fluent speaker review it, especially in a lower-resource language.

None of this is a reason to avoid prompting outside English. It's a reason to be deliberate about which language you're typing in, which language you want back, what register it should carry, and to verify rather than assume when it counts. What Is Prompt Engineering? and The Ultimate Prompt Engineering Cheat Sheet cover the broader technique set these instructions sit on top of, and Persona Prompting goes deeper on the register and voice questions raised above.