Why DeepSeek Answers in Chinese: Language Drift, Expert Mode, and How to Fix It

You're chatting in English. You ask a question. The reply comes back in Chinese. Nothing in your prompt was Chinese. No setting was changed. It just happens.

This is a real, documented issue with DeepSeek-V3. Users have reported it on the official GitHub repository. It's not a one-off glitch. The same English prompt can produce an English response one time and a Chinese response the next. There's no clear trigger.

The problem

The model sometimes switches languages mid-conversation. You write in English. It answers in Chinese. Or it mixes both. It can happen in a fresh chat. It can happen after many turns. It can happen in Instant Mode. It can happen in Expert Mode. There's no warning. There's no pattern you can easily predict.

Why does it happen?

The short answer: language confusion. It's a known problem in multilingual large language models. It has several causes.

Training data imbalance. LLMs learn from massive text corpora. If Chinese content is heavily represented in the training mix, the model develops a bias toward Chinese tokens. Research shows that the language distribution of training data acts as a confounder. It disadvantages languages that are underrepresented. DeepSeek is a Chinese-developed model. So Chinese data naturally carries significant weight.

Reasoning in Chinese. DeepSeek-V3 has a tendency to "think" in Chinese internally even when the final answer is supposed to be in English. The model's reasoning process—its chain of thought—may default to Chinese because that's more token-efficient for it. One commenter on the GitHub issue noted that telling the model to "translate to English" only limits its reasoning to Chinese while forcing the output into English. The internal language doesn't change.

Sampling temperature. Higher temperatures make the model more random. Research has found that language confusion is aggravated by high sampling temperatures. When the model has more freedom to pick tokens, it's more likely to slip into its dominant language.

Context pollution. Long conversations with mixed-language signals—search results, tool outputs, retrieved documents—can push the model toward an unintended language. Users have reported that the issue appears more often in longer sessions or when using DeepThink with Search enabled.

Expert Mode is not one model

Easy answer here. You're not interacting with just one DeepSeek model at once. You're interacting with many of them. At least two in Expert Mode.

One model replies in a similar way to Instant Mode. It works from pre-trained data. The second model activates when new data needs to be retrieved from the Internet. The searching model is still under heavy development. From time to time, something new may appear.

So you get more than just Chinese instead of English. You also get different AI personalities. One can deliver things not yet discovered. Instant Mode is even better at that because it interacts with end users much longer. The other limits replies to data it can read from the Internet.

After some time, Expert Mode should act more like Instant Mode. It should deliver both consistent replies and help develop things not yet known to humanity.

What can you do about it?

There's no guaranteed fix. The GitHub thread is clear on that. Even adding "Reply in English only" to the system prompt isn't guaranteed to work. But there are steps that reduce the odds.

Use a strong system prompt. Put a language instruction at the very top of your system message. Be explicit and repetitive. Something like: "Always respond in English. Do not switch to Chinese under any circumstances. Every response must be in English." This doesn't eliminate the problem, but it helps.

Lower the temperature. If you're using the API, drop the temperature to 0.3 or lower. Lower values make the model more deterministic. That reduces the chance of random language switching. Note that DeepSeek's API temperature scale differs from OpenAI's. A temperature of 1.0 on DeepSeek roughly equals 0.3 on OpenAI.

Start fresh. If the model switches mid-conversation, start a new chat. Long contexts accumulate language noise. A clean session with an explicit language instruction at the start is more reliable.

Turn off reasoning-heavy modes. Users have found that disabling DeepThink and resending the query often brings the response back to English. The reasoning mode is more prone to internal language switching.

Try Instant Mode if Expert Mode feels inconsistent. Expert Mode uses multiple models. That can add personality shifts and language drift. Instant Mode may feel more stable for simple English conversations.

For developers: intervene at inference time. Researchers have proposed methods like Language Neuron Intervention (LNI). They identify and manipulate language-specific neurons to steer output language. This is more complex, but it can be integrated into production systems.

The bigger picture: why bots reply in the wrong language

This isn't unique to DeepSeek. Language confusion is a systemic issue across multilingual LLMs. English-centric models are especially prone to it. But any model trained on mixed-language data can drift.

The core problem is that models don't have an explicit "output language" variable. They infer the language from context. When context is ambiguous—or when the model's internal reasoning pulls toward a different language—the output drifts.

English acts as a "semantic attractor" in many models. When drift happens, English is the most frequent fallback regardless of the intended language. But for DeepSeek, the attractor runs the other way. Chinese is the dominant internal language. So that's where it drifts.

Tips to keep bots on track:

• Be explicit about language in every system prompt. Don't assume the model will infer it.
• Keep conversations short and focused. Long contexts with mixed-language content increase drift risk.
• Avoid mixing languages in your prompts. If you write in English, don't paste Chinese text into the conversation.
• Test with different temperatures. Find the setting that works for your use case.
• Use post-processing as a safety net. A simple language detection check on the output can catch drift before it reaches users.
• Report persistent issues. The DeepSeek team is aware of the problem. User reports help prioritize fixes.

Language switching is annoying. It breaks trust. But understanding why it happens—and how to nudge the model back—makes it manageable.

References

1. DeepSeek-V3 GitHub Issue #1226: "[BUG] Model sometimes responds in Chinese even when the conversation is in English" — https://github.com/deepseek-ai/DeepSeek-V3/issues/1226

2. Stack Overflow: "Is there a way LLM answers only based on the context and also in the users asked language" — https://stackoverflow.com/questions/77548636

3. Xie, Y. et al. "Mitigating Language Confusion through Inference-time Intervention." COLING 2025. https://aclanthology.org/2025.coling-main.563.pdf

4. Marchisio, K. et al. "Understanding and Mitigating Language Confusion in LLMs." EMNLP 2024. https://aclanthology.org/2024.emnlp-main.380.pdf

5. Oh, J. et al. "OLA: Output Language Alignment in Code-Switched LLM Interactions." ACL 2026. https://aclanthology.org/2026.acl-long.2162.pdf

6. Kim, K. et al. "Query-Following vs Context-Anchoring: How LLMs Handle Cross-Turn Language Switching." MME 2026. https://aclanthology.org/2026.mme-main.13.pdf

7. arXiv: "Language drift patterns on the DuReader dataset." https://arxiv.org/pdf/2511.09984v1