Started seeing GPT 4.1 in eastus2 responding in spanish to a full english set of messages

Question

Started seeing GPT 4.1 in eastus2 responding in spanish to a full english set of messages

V 0

Is anyone else seeing an issue where GPT 4.1 in US East 2 has started randomly replying in Spanish for a series of system / human messages which are in English for a query condenser prompt (reformulate query based on a series of messages and latest user message).

The only thing we have in the system prompt is a:

"3. Preserve original language - Reformulate in the same language as the input. Never translate."

We saw one yesterday and thought it's a one off, but now got 2 more instances today!

SRILAKSHMI C 16,785 Reputation points Microsoft External Staff Moderator

2026-04-05T02:29:09.76+00:00
Hello V,

Thanks for reporting this,

I understand how confusing this can be, especially since your prompt explicitly enforces language preservation.

From what you’ve shared, this doesn’t look like a straightforward configuration issue. Since your inputs are fully in English and the behavior is intermittent, it’s more likely tied to model behavior in multilingual contexts rather than anything obviously misconfigured on your side.

What’s likely happening

GPT-4.1 is a multilingual model, and in certain edge cases it can:

Drift into another language (like Spanish) during reformulation/summarization tasks

Infer language from subtle or hidden signals in the conversation

Treat instructions like “preserve language” as soft guidance rather than a strict rule

This tends to happen more in:

Query condenser or rewriting scenarios

Prompts that are open-ended or not tightly constrained

Conversations with any prior multilingual traces

Why your current prompt isn’t fully preventing it

Your instruction “Preserve original language - Reformulate in the same language as the input. Never translate.” is correct, but in practice the model may still override it if:

It interprets the task as semantic transformation rather than strict formatting

There’s any ambiguity in output expectations

What to try next

1. Strengthen the language constraint

Instead of guidance, enforce a hard rule:

Always respond in English.

Add explicit output control

For example:

Respond ONLY in English.

Use structured output

For a query condenser use case:

{

This reduces ambiguity and helps anchor the response.

Add a post-response validation

Implement a lightweight check:

Detect response language

If not English - retry with stricter instruction

Please check your raw payload

As a sanity check:

Copy/paste the exact system + user messages

Ensure there’s no hidden Spanish text, metadata, or prior context leakage

Verify no cached conversation includes Spanish

Test a minimal reproduction

Strip everything down to:

One system prompt (with strict language rule)

One user message

If it still returns Spanish, that helps isolate it from your application logic.

Try another region or fresh deployment

Test the same prompt in another region e.g., West US

Or create a new GPT-4.1 deployment in East US 2

This helps determine if it’s region-specific or deployment-specific.

Please refer this

Azure OpenAI Service models (GPT-4.1 series): https://learn.microsoft.com/azure/ai-foundry/openai/concepts/models?tabs=global-standard-aoai#GPT-41-series

I Hope this helps. Do let me know if you have any further queries.

Thank you!
SRILAKSHMI C 16,785 Reputation points Microsoft External Staff Moderator

2026-04-06T12:25:51.1933333+00:00

Hello V,

Did you get any chance to review the above response. Do let me know if you have any further queries.

Thank you!
SRILAKSHMI C 16,785 Reputation points Microsoft External Staff Moderator

2026-04-07T12:37:14.0833333+00:00

Hi V,

Following up to see if the above answer was helpful. If this answers your query, please do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.

Thank you!

1 answer

Your answer

SRILAKSHMI C 16,785 Reputation points Microsoft External Staff Moderator

2026-04-06T12:25:51.1933333+00:00

Hello V,

Did you get any chance to review the above response. Do let me know if you have any further queries.

Thank you!
SRILAKSHMI C 16,785 Reputation points Microsoft External Staff Moderator

2026-04-07T12:37:14.0833333+00:00

Hi V,

Following up to see if the above answer was helpful. If this answers your query, please do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.

Thank you!

Answer 1

GPT-4.1 and similar LLMs can occasionally ignore or override language instructions in the prompt because responses are probabilistic and the model can be influenced by subtle signals in the conversation history or training data. Even when all visible messages are in English and the system prompt says “Preserve original language… Never translate,” the model can still:

Misinterpret the “original language” when there are prior multilingual tokens, names, or examples in the thread.
Drift to another language due to stochastic behavior, especially across multiple turns.
Be affected by other hidden or upstream instructions in the calling application or orchestration layer.

To reduce this behavior for a query condenser prompt:

Strengthen and repeat the language constraint
- Add explicit, high‑priority instructions such as:
  - “All input messages are in English. Always respond in English only. Do not use any other language under any circumstances.”
- Repeat this instruction near the end of the system message so it is more salient.
Make the task and language requirement more specific
- Instead of only “Preserve original language,” use something like:
  - “Reformulate the user’s latest query in English, preserving meaning and intent. Do not translate to or from any other language. If any non‑English text appears in the input, still respond in English.”
Avoid ambiguous or multilingual concatenation
- Ensure the application does not concatenate user data or prior messages in a way that mixes languages in a single instruction block. Guidance for multilingual prompts recommends avoiding concatenating user data directly into the prompt because it can create a multilingual direct instruction and confuse the model.
Validate and monitor
- Because model behavior is probabilistic and can vary for the same prompt, test the prompt across multiple runs and conversation histories.
- Implement guardrails in the application layer (for example, checking the output language and re‑prompting or correcting if it is not English).

These steps align with general prompt‑engineering and localization guidance: be explicit about language, avoid mixed‑language instructions, and tailor prompts for the specific task and locale behavior.

References:

Share via

Started seeing GPT 4.1 in eastus2 responding in spanish to a full english set of messages

1 answer

Your answer