Ask an AI chatbot the same ethical question in English and Chinese, and you might get two different answers.
Researchers built C-Voices, a dataset of 86,400 dilemma-based prompts across six languages, each pairing an action aligned with Chinese Social Values against a conflicting alternative. The 12 values span national, societal, and personal levels of Chinese culture. Testing several LLMs on identical dilemmas across languages, the team found that value preferences are model-dependent and language-sensitive - the same moral fork in the road produced divergent answers depending on which language it was asked in. To address this, the researchers also built a fine-tuning-free steering method that reads hidden-state differences between value-aligned and value-conflicting responses, then nudges a model's output toward a chosen value during inference.
The practical part is the transfer trick: a value vector extracted in one language can be applied to steer a model's behavior in a different language, without retraining. That means a company could calibrate a model's values once and expect the adjustment to travel across languages, a shortcut that sidesteps the inconsistency the researchers documented in the first place. The method also generalized to two existing value benchmarks, FLAMES and ValuePrism, suggesting the approach isn't just a one-dataset trick.
Translation, it turns out, isn't just a language problem for AI models - it's a values problem too.