Prompting in Dutch vs. English: When to Choose Which Language?

When developing AI applications for the Dutch-speaking market, developers and prompt engineers almost immediately run into the same question: should we feed the instructions to the model in English or in Dutch? Although modern Large Language Models (LLMs) such as GPT-4o, Claude 3.5 Sonnet, and Llama 3 are trained multilingually, the distribution of training data is far from equal. In most cases, English accounts for more than eighty percent of the total dataset.

This asymmetry in training data has direct consequences for the processing capacity, reasoning power, and efficiency of your prompts. In this article, we analyze the technical foundation of multilingual prompting, look at the power of mixed prompts, address the nuances of Dutch tone control, and offer a systematic testing approach to determine the best setup for your use case.

The Underlying Dynamics: Tokenization and Data Density

To understand why a model responds differently to an English prompt than to a Dutch one, we need to look at how LLMs process text. Models do not read words, but tokens (text fragments). The tokenizers of top models are optimized for the most common patterns in their training data. Because the dataset consists primarily of English, English words are often a single token. Due to their less dense representation, Dutch words are more frequently split into multiple sub-tokens.

Take the word geautomatiseerd. In English, the direct equivalent is automated. While the English word can be processed as a single token, the tokenizer often breaks the Dutch word down into three or four separate fragments (e.g., ge, auto, mati, seerd). This leads to two concrete effects:

• Higher latency and costs: More tokens mean higher costs per processed call and slower processing times for both input and generation.
• Loss of contextual density: Because the context window is filled with more individual tokens for the same content, the model has less 'effective space' for complex instructions or long documents.

Furthermore, the abstract reasoning patterns, logical chains, and programming instructions in the neural network are more strongly anchored in the English-language weights. This means that for complex logical tasks or structured data extraction, instructions in English often lead to a more stable outcome.

When Do You Choose English and When Dutch?

Determining the right language depends on the goal of the task. We divide tasks into two main categories: logical/technical tasks and cultural/language-bound tasks.

For advanced applications, we recommend experimenting with general prompt techniques such as role-prompting and chain-of-thought, where the language effect is even more pronounced.

Mixed Prompts: English Instruction, Dutch Output

One of the most effective patterns in professional prompt engineering for the Dutch language area is the hybrid prompt or mixed prompt. With this approach, you write the system role, instructions, constraints, and expected format in English, but explicitly specify that the final content must be generated in Dutch.

LLMs are highly capable of processing instructions in one language and delivering the result in another. This leverages the best of both worlds: the sharp instruction following of the English conceptual network and the desired Dutch output for the end user.

Example of a Hybrid Prompt Setup:

[SYSTEM INSTRUCTION]
You are an expert customer service reviewer for a Dutch e-commerce platform.
Analyze the provided customer complaint and generate a professional response.

[CONSTRAINTS]
- The output response MUST be written in fluent, natural Dutch.
- Tone: Empathetic, professional, and solution-oriented.
- Address the customer with 'u' (formal).
- Structure: 1. Acknowledgment, 2. Explanation, 3. Proposed solution.

[INPUT DATA]
Customer message: "Mijn pakketje is nog steeds niet aangekomen, terwijl er stond dat het er gisteren zou zijn!"

[OUTPUT FORMAT]
Return a JSON object with two keys:
1. "sentiment": (string) positive/neutral/negative
2. "reply": (string) The Dutch response text

By keeping the structure and logic in English, the model stays strictly within the boundaries of the JSON specification. With fully Dutch instructions, models sometimes tend to deviate from the JSON structure for these types of tasks by adding extra explanatory text.

Guiding Tone and Formality in Dutch (Je vs. U)

Dutch has a crucial social distinction that English lacks: the difference between 'je/jij' (informeel) and 'u' (formeel). LLMs, which naturally lean towards the most probable patterns, often tend to produce a hybrid form in generic Dutch texts where 'je' and 'u' are mixed. This looks unprofessional.

To guarantee the desired form of address, it is often not enough to simply say: "Write in the u-form". You need to anchor the context with examples. This is where few-shot prompting plays an indispensable role.

Guidelines for Formality Control:

• Explicit rules: Do not just define the pronouns; possessive pronouns like 'uw' and 'jouw' must also be explicitly mentioned.
• Provide examples: Include at least one example of a correctly formulated sentence in the target tone within your prompt.
• System prompt anchoring: Secure the tone in the system prompts of your application, so it is not overwritten by user variables.

Pitfalls of Translated Prompt Templates

Many teams start with a library of English prompts that have proven successful on the internet and translate them one-to-one into Dutch using Google Translate or DeepL. This regularly leads to poor performance. The reasons for this are:

1. Loss of jargon precision: English terms from IT or marketing such as "bullet points", "actionable insights", or "concise summary" can sometimes be difficult to translate into Dutch. If you translate "concise" as "beknopt", the model might cut content too rigidly in Dutch compared to the English original.
2. Long, complex sentence structures: Dutch grammar has different subordinate clause structures and verb orders than English. If a complex English prompt is translated directly, it results in long, complex Dutch sentences in which the core instruction for the model becomes blurred.
3. Unwanted cultural assumptions: Translated prompts often adopt American communication styles. This leads to Dutch that reads as overly enthusiastic ("Great news!", "We are incredibly excited to..."), which does not come across as authentic to a Dutch reader.

A Systematic Testing Approach: Measure the Difference Yourself

Since every model and task is unique, you should not make decisions about language preference based on gut feeling. Setting up a simple A/B test is the best way to determine the most optimal setup.

Step-by-Step Plan for Setting Up a Comparison Test:

1. Compile a dataset: Collect at least 20 to 30 representative input questions or scenarios from your practice.
2. Design three prompt variants:
   • Variant A: Fully Dutch (Instruction + Output).
   • Variant B: Fully English (Instruction + Output, then translated to Dutch).
   • Variant C: Hybrid (English instruction + Dutch output).
3. Evaluate based on fixed criteria:
   • Instruction Following: Is the JSON or structure 100% correct?
   • Language Quality: Does the Dutch text read naturally and correctly?
   • Tone Consistency: Is the u/je form used consistently?
   • Token Efficiency: How many tokens were used and what is the latency?

When working with external data sources or enterprise search systems, it is wise to check how the choice of language works in combination with your retrieval pipeline. Read more about this on the page about RAG for beginners on our educational subdomain.

Conclusion

Choosing the right language in your prompts is a strategic decision that directly impacts the quality, cost, and speed of your AI application. Use English for the control, logic, and structure of the model, and target Dutch for the final representation to the user. By maintaining this separation and testing your specific use case, you will build more stable and effective LLM applications.