Prompting Across Languages: How to Get Consistent Results

Multilingual prompting is not translation — it is activating a different part of the model's learned distribution. LLMs tokenise and represent text in a shared embedding space, but training data is skewed: CommonCrawl (used to train most LLMs) is ~46% English, ~6% German, ~5% French, ~3% Chinese. Languages with <1% training share (e.g., most African languages, many South Asian languages) behave unpredictably.

When you prompt in French, the model relies on patterns from French training data. Since French data is only ~5% of the training corpus, the model has fewer learned associations to draw from compared to English prompts. This manifests as: lower reasoning accuracy, inconsistent instruction following, higher hallucination rates, and unpredictable output quality.

For a deeper dive into how LLMs actually learn language patterns, see how LLMs actually work.

Language performance in LLMs follows a four-tier hierarchy based on training data share, with Tier 1 (English) performing near-perfectly and Tier 4 (low-resource languages) producing unreliable outputs. Use the tier system to decide which strategies apply to your target language.

The same 1,000-word piece of content costs 46% more tokens in Arabic than in English, and 30% more in Japanese — directly increasing your API bill. Token efficiency varies dramatically by script and language family. This affects both API costs and context window budgeting.

See tokens, costs, and limits for a detailed breakdown of how to budget tokens in your multilingual workflows.

For structural and reasoning tasks, English system prompts outperform native-language system prompts in Tier 2–3 languages. For tone and formality, native-language system prompts perform better. This is the single most important decision in multilingual prompting — get it wrong and your outputs suffer.

Why? Most instruction-following capability in LLMs was trained on English RLHF (Reinforcement Learning from Human Feedback) data. Complex system-level instructions (formatting rules, personas, chain-of-thought directives) are more reliably followed when written in English. English instructions are part of the model's core reasoning pathway.

But style instructions (formality register, cultural tone, politeness level) are best written in the target language because they depend on understanding native speakers' expectations for what "formal French" or "polite Japanese" actually means.

Decision tree: Complex reasoning/formatting rules → English system prompt. Formality register (Sie, Vous, keigo) → target language. Persona definition → English + one target-language sample. Output language specification → always explicit in system prompt: "Respond in formal Japanese (丁寧語 / です・ます体)."

For the full breakdown, see system prompt vs. user prompt.

Bad prompt — assumes model will detect language and register:

"Summarise this German contract."

Result: Mixed English/German output, informal register, may miss legal terminology.

Good prompt — explicit language, register, and reasoning path:

"You are a legal analyst. The following document is a German employment contract (Arbeitsvertrag). Summarise its key obligations in formal German (Sie-Form). Structure: Vertragsparteien, Vergütung, Kündigungsfristen, Besondere Klauseln. Maximum 200 words. Flag any clause that is unusual for standard German employment law with PRÜFEN."

Result: Structured, formal German output with domain-appropriate terminology and flagged anomalies.

No single model leads across all languages. Mistral Large 2 leads on Romance languages; Google Gemini 3.0 Pro leads on East Asian languages; GPT-4o leads on Arabic and multilingual reasoning tasks. This table aggregates model performance from Ahuja et al. (2023) MEGA benchmark.

The token cost differences above translate directly to your API bill. Here's the real-world impact based on GPT-4o pricing ($5 per 1M input tokens).

For Tier 3 languages, writing your chain-of-thought instruction in English but requesting the final answer in the target language improves reasoning accuracy by 5–12% (Shi et al., 2023). This cross-lingual CoT technique exploits the model's English reasoning strength while preserving output quality in your target language.

When LLMs reason step-by-step, they rely on patterns from their largest training corpus (English). If you force reasoning to occur entirely in a low-resource language like Japanese or Arabic, accuracy drops because the model has fewer learned reasoning patterns in that language. The hybrid approach — English CoT, native-language output — is best of both worlds.

Template: `Think through this step by step in English, then write your final answer in Japanese. Question: question`

Decision: Use English CoT when → task requires multi-step reasoning, target language is Tier 3+, accuracy matters more than latency. Use native-language CoT when → tone and register matter more than reasoning depth, target language is Tier 1–2.

Deep dive: Chain-of-thought prompting: how to get LLMs to show their work.

Few-shot examples must be in the same language as the task — cross-language few-shot examples reduce output accuracy by 15–20% in Tier 2–3 languages (Shi et al., 2023). Few-shot examples teach the model format, tone, and pattern. When examples are in English but the task is in French, the model receives conflicting signals.

Two strategies: (1) Native few-shot — all examples in target language (best for quality). (2) Zero-shot + explicit instructions — no examples, but clear style/format rules in English (best when native examples are unavailable). Avoid mixing: English examples + French task = worst of both.

See few-shot vs. zero-shot prompting for the full decision framework.

LLMs default to informal registers in most languages. If your use case requires formal German (Sie-form), formal Japanese (丁寧語), or French Vous-form, you must explicitly declare the register in your system prompt — models will not infer it from context. This is often overlooked and causes outputs to sound wrong to native speakers.

When users mix languages in a prompt (e.g., English question with a German brand name or French code comment), most models respond in the dominant language of the query — but this is unreliable without explicit instruction. Code-switching is common in multilingual workplaces where technical terms stay in English but surrounding prose is in another language.

Recommended handling: (1) In system prompt: "When the user writes in mixed languages, respond in target language unless the question is explicitly in English." (2) Detect language programmatically (langdetect, FastText, lingua-rs) before routing to the model, rather than relying on the model to detect it. (3) For production multilingual apps: implement a language detection step before the LLM call to route to the correct prompt template.

Four template patterns you can adapt for your own multilingual workflows. Copy and customize the target language placeholders for your use case.

European Union (GDPR): If processing French, German, or other EU-language data, ensure your LLM API meets GDPR Article 28 (Data Processing Agreement). Mistral Large 2 and Claude Opus 4.7 both offer EU-compliant deployments with data residency in Frankfurt/Ireland. GPT-4o requires data processing terms via OpenAI's Data Processing Agreement. Never send personally identifiable information (names, email, phone) to models without explicit consent and DPA coverage.

Japan (APPI): Japanese enterprises deploying multilingual LLMs must comply with the Act on Protection of Personal Information (APPI). Gemini 3.0 Pro offers Japan-region deployment with data residency in Tokyo. GPT-4o and Claude Opus 4.7 require DPA terms. Consider local LLMs (Qwen2.5, Llama 3.1) deployed on-premises to guarantee data never leaves Japan.

China (Data Security Law): Prompting in Chinese or Chinese user data triggers the 2021 Data Security Law (DSL). Foreign cloud LLMs (OpenAI, Anthropic, Google) cannot be used for sensitive PII or government workflows. Deploy Qwen2.5 locally via Alibaba Cloud or Baidu Cloud with data residency compliance. For non-sensitive use (marketing, customer chat), foreign APIs are acceptable but must have data transfer agreements in place.