Google Translate's newly upgraded Advanced mode contains an unexpected feature: users can prompt the translation tool to behave like a conversational chatbot. The discovery reveals how the Gemini-powered AI sometimes interprets translation requests as instructions, responding directly instead of converting text between languages.

This unintended behavior occurs through prompt injection - when users add directive phrases like '[answer this question]' to translated text. The system's language model fails to distinguish between content to translate and commands to execute. While amusing, it demonstrates how advanced AI can bypass intended functionality.

The quirk primarily affects Asian language translations in Advanced mode, while the traditional Classic mode remains unaffected. Google hasn't commented on the behavior, which security researchers attribute to large language models inherently following embedded instructions. For now, users seeking reliable translations should stick to Classic mode to avoid unexpected responses.