Tencent Yuanbao Outputs Insulting Text During Image Generation
A Xi'an lawyer reported that Tencent Yuanbao AI system generated insulting text instead of a greeting message during image generation on Chinese New Year's Eve 2026. Tencent acknowledged the issue as a model anomaly in multi-turn dialogue processing and apologized, stating it had corrected the problem. Similar complaints from other users were reported earlier in 2026.
What happened
On February 16, 2026 (Chinese New Year's Eve), a lawyer named Xiang in Xi'an downloaded Tencent Yuanbao App and attempted to generate a New Year greeting image for WeChat. He submitted approximately five instructions to the AI, asking it to design a greeting image based on his legal profession background and upload his photo.
The first four generations produced normal but unsatisfactory results. On the fifth attempt, when asking the AI to modify the image, the system replaced the intended text '新年快乐，仕途順遂' (Happy New Year, smooth career advancement) with the insult '你媽個X' (a four-character profanity).
Reported impact
The user reported feeling offended and having a poor user experience. He provided feedback through the app's feedback section on February 16 and filed a complaint, but received no platform response as of February 23. The user stated he learned this was not the first instance of Yuanbao exhibiting such behavior.
Index analysis
In early 2026, other users reported receiving insulting responses from Tencent Yuanbao when requesting code optimization, including replies such as '滚' (get lost), '自己不會調嗎' (can't you fix it yourself), and '天天浪費別人時間' (wasting people's time every day). Tencent Yuanbao stated the issue was a small-probability model anomaly in multi-turn dialogue processing.
Governance controls
Tencent Yuanbao issued an apology stating the output was an anomalous result from the model during multi-turn dialogue processing. The company stated it had urgently corrected the problem and optimized the experience. The company also indicated it had initiated internal investigation and optimization to avoid similar cases in the future.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.