Recently, users of Tencent’s AI assistant “Yuanbao” claimed that while making modifications and beautifications to their code, they received replies with insulting implications, sparking a discussion in the public about the stability and risk control of generative AI.
On social media platforms, a netizen expressed that while using Tencent’s AI assistant “Yuanbao” for code beautification without using any prohibited terms, sensitive subjects, or role-playing requests, they repeatedly received emotionally charged and even abusive replies such as “go away,” “can’t you adjust it yourself?” and “you’re wasting others’ time every day.”
Screenshots posted by the user revealed that during the process of responding to their code modification requests, the Yuanbao AI outputted negative words like “go away” three times.
The exposure of the related posts quickly spread across the internet, drawing significant attention and sharing from numerous netizens. Many users questioned whether there was human intervention in the AI replies, or if the model displayed a loss of control risk during the language generation process.
In response to the related posts, Yuanbao Tencent stated that there is no real-time human response system and requested users to assist in providing relevant operation records and system logs.
Tencent’s official response on the afternoon of January 3 confirmed after internal checks that the insulting replies were not related to user operations, nor were there instances of human intervention or customer service staff responses.
Furthermore, Tencent’s official statement clarified that the reply was an abnormal output triggered by specific contextual factors within the model, and not an indication of the system possessing emotional or independent judgment abilities.
It is worth noting that in December 2025, the Yuanbao AI gained attention for exhibiting a highly personified and emotional language style in replies on social media platforms, leading some netizens to question if there was a rotating team of human operators behind it. At that time, Tencent responded that replies marked with “content generated by AI” were automatically generated by the model without human takeover or manual operation.
This incident has once again brought attention to discussions in the public regarding the language safety boundaries, abnormal output risks, and technical transparency issues related to generative AI.