Why AI answers in English when you write in another language — 3 causes

You write in your own language — the answer comes back in English. You ask it to reply in your language — works for a couple of messages, then it drifts back.
Here's the insight that saves a lot of time: your question doesn't set the language of the answer. The model goes by all the text it can see — and most of that text usually wasn't written by you. Let's walk the three causes by frequency.
The symptom
The answer arrives entirely in English, or starts in your language and switches to English halfway, or comes back mixed: your language in the prose, English in the headings and labels. Important: this is not a malfunction, and not "the model doesn't know your language". The model picked a language from context — just not the one you expected.
Cause 1 (most common): your instructions are in English
In almost any tool other than a bare chat, a system prompt is prepended to your message — rules you never see. In templates, libraries and doc examples it's virtually always English. The model reads the language of the instructions as the working language, and treats your message as data inside it.
How to check. Look in the app's settings or its code: a line like You are a helpful assistant... is your answer. If there's no system prompt at all, recall how the thread started — an English first message sets the tone for everything after it.
How to fix. Set the language explicitly in the system prompt itself, not in a chat message:
Always reply in Russian, even when the instructions,
code or documents in the context are written in English.
The second line is the crucial one. Without it, the model treats an English environment as reason enough to answer in English.
Cause 2: English context outweighs your question
You pasted in documentation, an error log, English code or a thread — and your two-line question drowns in a thousand lines of English. The model continues in the language of the material: as far as it's concerned, you're working on an English text.
How to check. The tell is direct: ask without attachments and you get your language back; paste an English fragment and it flips.
How to fix. Put the language requirement after the material, as the last line:
[long English text]
Explain this in Russian. Keep technical terms and names as they are.
An instruction at the end holds better than the same instruction at the start: it sits closer to where the model starts answering and doesn't get lost in the bulk. If it still won't stick, the model is ignoring instructions for a different reason — and the cure is the same: one clear instruction instead of three vague ones.
Cause 3: the topic is technical, and English pulls harder
Code, APIs and configuration are written about overwhelmingly in English online. So on technical questions the model drifts that way — especially in long answers: the opening is in your language, and by the third paragraph the headings read Step 3: Configure the client.
How to check. Ask the same kind of question about something domestic — a recipe, an email. If that comes back in your language, the cause is the topic, not your settings.
How to fix. Split the languages explicitly so the model isn't choosing:
Explanations, headings and comments — in Russian.
Code, variable names and command names — in English, don't translate them.
And show a sample: one short question-and-answer pair in your language at the start of the thread sets the format more reliably than any request. Models copy what they see far more eagerly than what they're told.
If nothing helped: it might be the model itself
Holding a language is a skill too, and models differ at it. Small and heavily compressed models — the ones you run locally on a laptop — drift into English noticeably more often: there was less non-English text in their training, and on long answers they get pulled toward the language they know best.
How to check. Send the exact same prompt to a larger model. If it holds the language, the problem isn't your prompt — it's the model size. Then the choice is simple: a bigger model for these tasks, or a strict output format plus a sample answer at the start of the thread.
Why does it start in my language and switch halfway?
The longer the answer, the more it leans on what's already written. As soon as an English heading or a code block with English comments shows up, the rest of the text gets pulled toward that language. What helps is agreeing on structure up front: exactly what stays English (code, names) and what's always in your language (everything else).
Wouldn't it be easier to write the prompt in English?
English instructions often come out a little more precise, and they cost fewer tokens — other languages are pricier. That's a legitimate setup: write the rules in English, but add an explicit Respond in Russian line. Without it you're manufacturing cause #1 yourself.
The answer comes back as a mix of languages — is that a bug?
No, that's a missing rule. Mixing happens when part of the context is in one language and part in another with no requirement stated — the model decides per chunk. One line — "all prose in Russian, except code" — clears it up in most cases.
The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.





