A new study has recently discovered that the language you use to interact with an AI chatbot can influence your answers. According to a new study by Nature, popular AI assistants, including ChatGPT and Claude, gave very different responses to the questions asked about China. This was observed when the prompts were written in Chinese instead of English. It was observed that the responses generated were more favourable and less critical when the questions were asked about China in Chinese. This finding has started discussions around the bias, multi-lingual data training and how large models behave across languages. What Did the Study Find?The researchers largely focused on AI responses to politically sensitive questions related to China in both English and simple Chinese. Here are the findings:The same question sometimes produced different answers depending on the language usedChinese-language responses were, in most cases, more aligned towards favourable narratives about ChinaEnglish responses included critical viewpoints and had a wider range of perspectivesThe differences seen in both ChatGPT and Claude, though both are developed by US-based companiesThe researchers believe that these differences may come from variations in multilingual training data, language-specific safety systems and reinforcement learning process instead of a deliberate change in behaviour. Why Do AI Responses Change Across Languages?The systems are largely language-trained on huge datasets from books, websites and other publicly available content written in multiple languages. However, this source data is usually not perfectly balanced. Factors that can influence AI responses may be:Different sources available in each languageVariations in the cultural context of the global languageLanguage-specific moderation systemsSafety policies applied to multilingual contentReinforcement learning based on language-specific examplesThis is why not everything remains the same when the prompts are entered in different languages. Why This Matters?The study highlighted that despite AI tools becoming increasingly popular, they are not meant to be trusted blindly. If the AI responses vary depending on the language, users may unknowingly receive different responses for the same topic. This means that there is a need of a greater transparency in how AI models are trained across different languages. Should Users Be Concerned?In most cases, this should not be a matter of concern. But it is always recommended to treat AI-generated responses as informational instead of blindly trusting them. And if you are researching topics related to politics, history or international relations, you should:Verify information using multiple trusted sourcesCompare responses across different AI models if accuracy is very important for youCheck original reports, academic research and official statements before drawing conclusionsCheck the sources of the informationThis approach can help in preventing bias and helping you with accurate information. Other Frequently Asked Questions (FAQs) That You Should Know:Why did ChatGPT and Claude give different answers in English and Chinese?The study suggested that the differences usually arise due to multilingual training data, language-specific safety systems and the variations in the information available in different languages. When the source of information differs, the output will be different. Does this mean AI models are biased?Nothing can be clearly stated as of now because the differences are usually seen on sensitive topics and across languages. Researchers currently say that more work is needed to understand this better. Should I trust AI for political or historical information?AI can be a good starting point to help you collect the data. But it is important to verify the sources and check for multiple results before you jump to any conclusions.end of article