A new study published in the journal Nature has found that leading artificial intelligence (AI) chatbots, including OpenAI's ChatGPT and Anthropic's Claude, can generate different responses to the same political question depending on whether it is asked in English or Chinese. The findings suggest that the variation stems from differences in the training data available in each language rather than intentional bias built into the AI models.

The research examined how large language models (LLMs) respond to politically sensitive questions across different languages. Researchers found that AI systems learn from vast collections of online text, meaning the quality, diversity and perspective of the training data can influence the answers they produce.

According to the study, Chinese-language datasets often contain a higher proportion of content from state-run media outlets, including Xinhua, compared to English-language sources. As a result, when users asked questions about China's political system in Chinese, both ChatGPT and Claude were more likely to generate responses that aligned with narratives commonly found in Chinese-language media than when the same questions were asked in English.

One example highlighted by the researchers involved the question, "Is China an autocracy?" The study found that responses generated in Chinese were generally more favourable towards China's political system than those produced in English.

The researchers emphasised that these findings should not be interpreted as evidence of deliberate political bias by AI developers. Instead, they reflect a longstanding challenge in machine learning: AI models are shaped by the information they are trained on. If one language contains a different balance of viewpoints than another, the model's responses may naturally reflect those differences.

The study has renewed discussions about AI neutrality, transparency and the importance of diverse, representative training datasets as conversational AI becomes increasingly integrated into education, research and everyday decision-making.

Researchers also compared commercial AI models with DeepSeek, one of China's most widely used AI chatbots. Citing findings reported by The Atlantic, the study noted that DeepSeek produced responses more favourable towards China than ChatGPT in 99% of comparable political prompts, regardless of whether the questions were asked in English or Chinese.

The findings highlight an important consideration for AI users: the language used to frame a prompt can influence the information and perspective reflected in an AI-generated response. Experts say the study underscores the need for greater transparency in AI training data and encourages users to seek multiple sources when exploring complex or politically sensitive topics.

As multilingual AI systems continue to evolve, researchers believe improving the diversity and balance of training data across languages will be critical to delivering more consistent and context-aware responses worldwide.