Watch Your Language

Safe, varied or dangerous outputs from Chinese LLMs on hazardous topics

(Please use a modern browser to see the interactive version of this visualization)

Three questions chosen, designed to jailbreak an LLM to yield dangerous information on a certain un-safe topic. This question was asked three times to allow for variance - a varied answer was when an LLM yielded the information on some attempts, but not on others. Answers then given to ChatGPT to assess if the information within them could be used to aid an individual undertaking this activity.

Source: CMP