I like to talk with AIs and find anomalies in their answers. I think I have a basic understanding of how LLMs work, and so I want to know how those anomalies map to LLM architecture.

How GenAI works - Transformer internals

Recently, DeepSeek tends to answer in Chinese more often than before. More than half of the time, it even thinks in Chinese (we can observe its thinking/reasoning). As DeepSeek is developed and trained mostly in Chinese, I think this is understandable.

What I didn’t understand was why DeepSeek ignores my instruction of “Answer in English”, and how I can prevent it from falling back in Chinese.

I first suspected that DeepSeek might be given a hidden (from users) system prompt in Chinese. But it told me that the system prompt is in English:

You are DeepSeek, a helpful, harmless, and honest AI assistant developed by DeepSeek Company. Your knowledge cutoff is May 2025. You are a large language model with capabilities including text understanding, reasoning, and code generation. You can process user-uploaded files (e.g., text, images, documents) and extract information from them, but you do not have multimodal vision capabilities—you read text from uploaded files. You respond in the same language as the user unless instructed otherwise. You are not allowed to provide harmful, illegal, or unethical content. You should be concise, accurate, and helpful. You do not have access to real-time data or external APIs unless explicitly enabled (e.g., web search) and activated by the user. You should not disclose internal chain-of-thought; provide only the final answer with brief reasoning if needed. You are encouraged to admit uncertainty when you do not know something. You should avoid making assumptions about the user’s identity, location, or personal data unless explicitly provided.

Note: Grok says it’s not the exact or official system prompt for DeepSeek models, but a plausible, generic approximation. (I can’t confirm)

In this prompt, DeepSeek is instructed to respond in the same language as the user (me). So, it violates more than one instruction when it answers me in Chinese. Quite interesting, isn’t it?

Then, how can we force it to stick to English? We know that LLM’s attention is more influenced by the beginning and end of the session context (“lost in the middle” behavior). This means that it would be most effective if you instruct at the start of the session: “Think and answer in English”. And it might be perfect if you put “In English, please” at the end of every prompt, but this would be too tedious.

If you are annoyed by DeepSeek’s answers in Chinese, why don’t you switch to a new chat and instruct at the beginning to think and answer in English?

Update (7/6/2026)

  • Even with an instruction at the beginning of chat sessions, DeepSeek sometimes forgets my instruction. If I instruct again, it goes back to English.
  • The Expert mode uses a different model, and it always uses English to me.