Researchers tested hundreds of questions in different languages and response formats. Simplified Chinese prompts produced more pro-government responses for most models, while some non-Chinese models also sometimes echoed official positions. Political tendencies were harder to detect in realistic work tasks.
Vals reported that targeted fine-tuning reduced such responses for Thomson 1.0 Small from 84.6% in its Qwen3.6 35B base model to 15.0%. The company says its questions are not typical user prompts; refusals counted as neutral, and some models were accessed through third-party providers. These limits matter when interpreting the results.
Why it matters
Model behavior may depend on language and context, and the study’s alignment score is not a measure of truthfulness.
Something to try
If comparing models, try asking a sensitive question in more than one language and compare the open-ended answers.
Start the conversation.