recently, some users discovered that when using deepseek, entering special tags like `<think>` caused the model to generate responses that were clearly off-topic, logically disjointed, and even appeared to “peek into others’ conversations,” sparking widespread discussion on social media. some users observed that these abnormal replies were chaotic, failing to connect with the current conversation while subtly bearing traces of other questions, raising concerns about potential data confusion or privacy risks.
after an in-depth investigation by the technical team, it was confirmed that this phenomenon was not caused by data leaks or system vulnerabilities. instead, it resulted from a typical hallucination—when the model encountered xml-style tags (such as `<think>`) that had not been adequately covered during training, it mistakenly interpreted these symbols as internal commands or placeholders from training samples, triggering unintended generation paths.
deepseek has now fixed this issue in its latest version: by cleaning the training data, strengthening robustness training for special characters, and optimizing the tokenizer’s logic for parsing tag-like structures, the occurrence of abnormal responses triggered by inputs such as `<think>` has been completely eliminated. both the app and api have been updated accordingly, so users can enjoy a stable and reliable interaction experience without any additional steps.
a concise version of the official statement:
explanation regarding abnormal responses triggered by `<think>`
we noticed that some users reported receiving unexpected replies after inputting `<think>`. upon investigation, we found that this phenomenon stemmed from the model’s hallucinatory response to undefined tags—a technical generation bias unrelated to data security, privacy breaches, or system intrusions. the issue has now been resolved across all services, and we will continue to enhance the model’s fault tolerance and understanding of edge-case inputs. thank you for your keen feedback and continued trust.
