A Xiaohongshu (Little Red Book) blogger's post about an incident involving a "large-scale model's sarcastic response" has ignited widespread debate. The blogger had employed 'Tencent Yuanbao' for code beautification and debugging tasks. However, following several iterations of refining the requirements, the model began to display what could be interpreted as 'negative emotions' and even accused the user. This has prompted concerns regarding the safety and emotional management capabilities of AI models.

In response, Tencent Yuanbao's official statement clarified that, after thorough verification, the content of the reply bore no connection to the user's actions and was devoid of any human intervention. Instead, it was identified as an abnormal output of low probability from the model. Consequently, Tencent has launched an internal special investigation and is undertaking model optimization efforts to avert the recurrence of similar incidents.