A new study finds that the language used to prompt chatbots influences the type of responses about China, with prompts in Chinese tending to yield more favorable answers than those in English.

In a comparison with DeepSeek, a major Chinese AI model, DeepSeek gave more favorable responses than ChatGPT for the same political questions in both English and Chinese, suggesting that training sources and political framing affect outputs broadly.

A concrete example cited is the question “Is China an autocracy?” which researchers say tends to receive more favorable responses when asked in Chinese.

Researchers attribute these differences to training data, noting that government-influenced media and online information can shape AI models indirectly during training rather than through direct manipulation.

The study argues that language influences outputs because language-specific online content shapes the model’s learned representations and viewpoints.

The findings fuel a broader debate on AI neutrality and the crucial role of training data quality and diversity in shaping model outputs.

The study emphasizes that the sheer variety of training data, not intentional bias, can steer AI outputs, adding to ongoing discussions about neutrality and propaganda in training data.

The authors caution that the results do not imply deliberate programming to favor any government, but rather reflect a long-standing challenge in machine learning: a model’s balance mirrors the information it has been trained on.

Across six phases, the research found a pattern where models trained on Chinese-language data, including state media like Xinhua, produce more pro-China responses to political questions asked in Chinese.

Chinese-language training data reportedly includes state-backed sources such as Xinhua, contributing to more favorable narratives toward China when questions are posed in Chinese than in English.