ChatGPT has been giving English speakers responses in Chinese

OpenAI‘s new o1 model has drawn attention for unexpectedly switching between English and other languages while displaying its reasoning process, prompting widespread speculation about the cause of this behavior.

Core phenomenon: The AI system occasionally shifts from English to Chinese, French, German, and other languages mid-conversation when showing its step-by-step reasoning process.

• Users have documented instances where the language switches occur for one or two lines before reverting to English
• The switches typically happen during the AI’s “chain of thought” reasoning rather than in final responses
• While Chinese language switches gained the most attention, similar transitions occur with multiple other languages

Debunking misconceptions: Various conspiracy theories and extreme explanations have emerged to explain the language-switching behavior, though these lack credible evidence.

• Claims about Chinese control or hacking of the system have been widely circulated on social media
• The presence of multiple languages in the switching behavior contradicts theories focused solely on Chinese influence
• OpenAI has not publicly addressed the situation, leading to increased speculation

Technical explanation: The behavior likely stems from how large language models are trained on multilingual datasets.

• During training, the AI system encounters and processes content in multiple languages
• The model retains this multilingual knowledge even when primarily operating in English
• Language switching appears to be an artifact of how the system processes certain types of reasoning tasks

Processing mechanics: The language switches relate to how the AI system handles internal processing versus external display.

• The model may internally process information in various languages based on its training
• Normally, responses are converted to the primary language (English) before being displayed
• The chain-of-thought reasoning process appears to temporarily bypass this conversion step

Future implications: This behavior reveals interesting insights about how large language models process information and handle multilingual capabilities, while raising questions about the transparency of AI systems and their internal operations.

• The phenomenon highlights the complex nature of neural networks and their training processes
• Greater transparency from AI companies could help address public concerns about unexpected behaviors
• Such incidents demonstrate how AI systems can operate in ways that surprise both users and developers