Gemini API bug report: English prompt and English sources produced a Chinese script
Submission title
[BUG] gemini-3-flash-preview unexpectedly returned Chinese dialogue for an English-only task context
Summary
On August 11, 2026, a production content-generation call to `gemini-3-flash-preview` received an English prompt, English source material, English research context, and an English JSON output contract. The returned dialogue content was almost entirely Simplified Chinese while retaining the requested JSON structure and speaker identifiers.
No Chinese source material or translation request was intentionally supplied. The application did not explicitly state “English only,” so this report does not claim violation of an explicit language constraint. It reports a severe and unexpected output-language switch despite the complete task context being English.
The unexpected output passed into text-to-speech and created a video containing only Chinese speech.
Expected behavior
Return a JSON array of dialogue written in natural American English, matching the English prompt, examples, source material, host descriptions, and requested audience.
Actual behavior
The model returned a correctly structured JSON dialogue array whose dialogue text was predominantly Simplified Chinese. A subsequent quality-review call to the same model also preserved Chinese rather than returning the script in English.
Example output excerpt:
HOST_A: 今天是二零二六年八月十一日。如果我们回顾一下昨天的国会记录……
HOST_B: 所谓的“九十票挑战”，听起来就像是参议院完成了一个复杂的跨学科小组作业……
The saved script contains 3,609 Han characters. The separately generated extended-format episode, using the same underlying English source corpus and the same text-generation model, contained zero Han characters and was generated in English.
API and runtime details
-
API: Gemini Developer API through the official Python SDK
-
Model:
gemini-3-flash-preview -
SDK: `google-genai 1.67.0`
-
Python: `3.11.6`
-
Operating system: macOS 26.2, build 25C56
-
Architecture: Apple Silicon / arm64
-
Call form:
client.models.generate_content(model=model, contents=prompt) -
Sampling configuration: no explicit temperature, top-p, top-k, or candidate-count overrides; model defaults were used
-
Tools/function calling: none
-
Input type: one text prompt
-
Requested output: JSON array of dialogue turns
-
Approximate script-generation input: 10,800 tokens / 42.2 KB
-
Approximate script-generation output: 3,700 tokens
-
Script-generation latency: 46.8 seconds
-
Quality-review call: approximately 1,200 input tokens and 1,300 output tokens; 34.6 seconds
Incident time
-
Application run began:
2026-08-11T21:32:42.776Z -
Unexpected script was written: approximately
2026-08-11T21:39:34Z -
Main run finished:
2026-08-11T22:06:44.321Z
Please correlate calls from the affected Google AI project during approximately 2026-08-11T21:38:00Z through 2026-08-11T21:40:30Z.
Request context
The prompt:
-
Was written in English.
-
Required JSON dialogue turns with speakers `A`, `B`, or `PAUSE`.
-
Did not request translation or mention Chinese, Mandarin, localization, or a non-English audience.
-
Did not contain an explicit sentence requiring English-only output.
The model preserved the requested structure but unexpectedly changed the dialogue language.
Reproduction information
The event was a single probabilistic generation and has not been intentionally rerun because the call is paid and the bad output already caused downstream production impact.
The application did not persist the exact assembled prompt or the SDK response object’s request/response identifier. Available evidence includes:
1. The prompt template used to assemble the request.
2. The complete unexpected Chinese response after JSON parsing.
3. The complete English extended response generated later from the same source corpus.
4. Pipeline logs containing model name, approximate token counts, latency, and downstream processing.