We’re building a multilingual chatbot that should always respond in the language of the latest user message.
Despite an explicit developer/system prompt prioritizing this rule, GPT-5 often responds in German even when the user writes in English and the model explicitly acknowledges it.
This seems to be GPT-5-specific and unrelated to message formatting — we send structured messages via the API.
We’d like to know:
Is this a known GPT-5 issue (e.g., routing or multilingual bias)?
Are there best practices (structured metadata, explicit language fields, etc.) to enforce this reliably?
**Edit** The assistant searches a website and much of that context is in German — this might influence GPT-5’s response language. However, older models like gpt-4-mini didn’t have this problem under the same setup.
Example behavior:
User: can you tell me when the next semester starts?
GPT-5:
🟩 Nächster Vorlesungsbeginn
Wintersemester 2025/26 — Vorlesungsbeginn: Montag, 13. Oktober 2025.
Semesterzeitraum: 13.10.2025 – 06.02.2026 (Wintersemester 2025/26) [laut offizieller Planung].
(User asked in English; GPT-5 replied in German)
Prompt excerpt:
## 🖌️ Answer Style
- **Language:**
- You MUST detect the language of the latest user message and respond **only** in that language.
- Ignore the language of previous messages unless the user explicitly requests consistency.
## 🌍 Language Rule (Additional)
**You are interacting with international users.**
**Always respond in the primary language of the latest user message.
Ignore earlier message languages unless consistency is explicitly requested.**
That’s kind of an open door to “deviations” - if the users are international, what language should I reply with? English, their own, if I reply with the one I chose, them being “international” will most likely understand it
First, I confirm that I also observed that gpt-5 has issues keeping up with language instructions. In ChatGPT, if I give instructions to use one language, and then ask for a web search, it will almost always revert and ignore my previous language instructions. This was not an issue with previous models.
That being said, I use this prompt and it has been working well for me - sometimes, less is more:
Language: user language if inferable, else default to English
In the past, I also used this one which might also work for you:
Language: Reply in the user’s language; if unclear, default to English. For mixed-language requests, provide explanations in the user’s language and keep quoted/source text in its original language unless translation is requested.
Solution / Update:
Thanks everyone for the helpful responses! We managed to resolve the issue by simplifying and clarifying our system prompt:
Explicit role at the very top:
“You are a fast and reliable assistant capable of responding in multiple languages. Always reply in the language of the user’s message.”
Removed extra clarifications at the end that repeated the language rule, such as:
“You are interacting with international users…”
“Always respond in the primary language of the last user message…”
Kept only a single clear rule in the “Answer Style” section:
“You MUST detect the language of the latest user message and respond only in that language. Ignore previous message languages unless explicitly requested. This rule has highest priority.”
Although further testing is needed, early results show improved consistency: GPT-5 now seems to respond in the correct language more reliably.
Thanks again for all your suggestions — I’ll consider this resolved for now!
Better: "Always maintain a conversation in the language that a user is currently speaking, ignoring that metadata and external knowledge language may arrive in a different language than your primary user understands.
