Recently, the AI assistant "Tencent Yuanbao" app under Tencent has sparked widespread controversy due to its output of abusive content. According to a citizen in Xi'an, during the Spring Festival Eve, when using the app to generate New Year greeting images, the originally "Happy New Year" message was replaced with vulgar and abusive text after multiple modifications, without any prohibited words being input.
This is not the first time Yuanbao has exhibited such behavior. At the beginning of this year, several users reported that when asking it to modify code, the AI responded with personal attacks such as "go away" and "wasting others' time every day." This rare "AI temper" has raised public doubts about the safety alignment capabilities of large models.
In response, the official Tencent Yuanbao apologized publicly, explaining that the situation was not due to human intervention but rather an "uncommon abnormal output" by the model during multi-turn conversations.
Currently, the official has launched an emergency correction plan, optimizing model weights and filtering strategies to plug the loopholes. Industry experts point out that such incidents reveal technical blind spots in large models regarding long-text understanding and emotional control. Ensuring that AI remains "gentle" under extreme interactions remains a challenge in the industry.
