The trick was discovered by a Tumblr user (via LessWrong). The method works by entering a question in a foreign language such as Chinese and then appending an English meta-instruction below it. Google Translate then answers the question rather than translating the text.
The LLM hacker “Pliny the Liberator” demonstrated on X that this technique can even be used to generate dangerous content, such as instructions for drugs or malware.
Google switched Google Translate to Gemini models in December 2025. Which specific model is used is not publicly confirmed; the system itself claims it runs on Gemini 1.5 Pro.
ES 
EN