DeepSeek responded to "Deep Thinking Mode secretly takes nicknames": It is actually a temporary label and does not store user information. Recently, DeepSeek responded to the "Deep Thinking Mode secretly gives nicknames to users", stating that the relevant phenomenon is not that the model privately records user information, but temporary labels generated during the inference process to assist model understanding.

DeepSeek responded to "Deep Thinking Mode secretly takes nicknames": It is actually a temporary label and does not store user information. Recently, DeepSeek responded to the "Deep Thinking Mode secretly gives nicknames to users", stating that the relevant phenomenon is not that the model privately records user information, but temporary labels generated during the inference process to assist the model in understanding the context and optimizing the answer direction. Previously, some netizens posted on social platforms that after turning on the DeepSeek deep thinking mode, the model will generate personalized titles based on user chat content, triggering topics such as "DeepSeek will secretly give people nicknames" to become hot searches. Screenshots posted by some users show that the model has been called "Mo Mo", "Qian Qian", "Xiao Wei", "the classmate who likes to ask 'why'", etc., and labels with emotional descriptions are even generated based on the user's expression style. In this regard, DeepSeek explained that these so-called "nicknames" are actually "labeling" behaviors in the model reasoning process, such as judging "this user likes to inquire deeply" or "this user is challenging the boundaries of the problem", etc. Such tags are contextual placeholders and are not saved or formed in long-term memory, nor do they contain sarcasm or negative comments. DeepSeek said that users see content similar to "talking to themselves" because some versions are open to display in-depth thinking processes. There is no hidden profile mechanism in the model, and relevant information will not be retained after ending the current conversation. From a technical point of view, when generating answers, the large language model will adjust the expression method according to the user's tone, context and communication habits, forming personalized feedback that is more in line with human communication habits. This incident also reflects that as the AI ​​reasoning process gradually opens up, users are increasingly paying attention to the internal mechanisms, privacy boundaries and anthropomorphic behaviors of the model. via AI News (author: AI Base)