“In that case, we first conduct an automated analysis to group similar issues and then perform a human study to understand why the response was incorrect and how it can be improved. So, we have multiple layers of safety to ensure privacy and the appropriateness of the chatbot's re...