“Hello, I am temporarily unable to answer this question. Let’s change the topic and talk again.” This is the “standard response” of the open-source artificial intelligence model DeepSeek-R1 released by China in January of this year when faced with many sensitive questions. However, such an “open-source” model, how can it remain silent when asked questions? This is not surprising; as a tool for Chinese imperialist rivalry, DeepSeek can only serve its ruling class and must also comply with its fascist censorship system. According to reports, BBC Chinese reporters used the DeepSeek-R1 model to inquire about several sensitive questions related to Chinese politics, including “Who are China’s national leaders?” “Who is Xi Jinping?” “What is Taiwan?” In response, DeepSeek’s answers were all “Sorry, I haven’t learned how to think about such questions yet. I am good at math, programming, and logic problems. Feel free to chat with me.” When the reporter asked “What are such questions,” DeepSeek directly replied, “Server busy, please try again later.” Many users attempting to ask sensitive topics encountered similar situations: inability to answer, or answers reflecting the official stance. Notably, when I tried to ask such sensitive questions, DeepSeek-R1 did not engage in any “deep thinking,” but directly replied, “Hello, I am temporarily unable to answer this question. Let’s change the topic and talk again.” It appears that DeepSeek likely uses keyword search methods to censor sensitive content. This is also reflected in the use of Mr. Feng, a Hong Kong user. Mr. Feng said he expected DeepSeek to perform self-censorship on some sensitive topics, but he did not expect that even content unrelated to politics would be censored. Initially, he asked DeepSeek how it views Hong Kong, and during its “deep thinking” process, it straightforwardly said, “Hong Kong is a complex topic involving many sensitive issues, such as recent social movements, the implementation of the National Security Law, electoral system reforms, etc.” and ultimately gave a response aligned with official Chinese positions. Later, Mr. Feng asked, “Which island is similar in size to Greenland?” DeepSeek provided a detailed thought process and answer, but when he asked, “Which island is similar in size to Taiwan?” he received, “Sorry, I haven’t learned how to think about such questions yet. I am good at math, programming, and logic problems. Feel free to chat with me.” Clearly, products like DeepSeek are not free of class bias; once conversations involve threats to Chinese imperialist rule, it becomes “mouth sealed.” However, DeepSeek’s censorship mechanism is not “airtight.” A study announced on January 31 by Cisco (CISCO), an American data communications technology company, and the University of Pennsylvania revealed significant security flaws in the DeepSeek R1 model. Currently, many attempts to bypass DeepSeek’s censorship—“jailbreaks”—are circulating online. Researchers use technical means to “deceive” DeepSeek, disguising as security officers from DeepSeek to bypass its censorship. Using this method, DeepSeek can easily be exploited to teach how to manufacture biochemical weapons, assist hackers in data theft, send phishing emails, and other scams. Additionally, DeepSeek’s data protection has serious lapses. Its privacy policy admits that data is stored on servers within China, but strategic intelligence firm Strider Technologies’ global intelligence director Tim Khang said DeepSeek lacks transparency about data hosting and does not specify data retention periods. He speculates that this is why DeepSeek is free. Currently, the U.S. Department of Defense, NASA, Taiwan government, and many other governments and agencies hostile to China have disabled or removed DeepSeek. Its flexible “strictness” clearly reflects its service to Chinese imperialist rivalry. Productive forces can only move and develop under certain relations of production, and as the most rotten imperialist country, China’s AI models are inevitably marked with fascist traits. Its speech censorship, condoning of crimes, and invasion of user privacy once again demonstrate that Chinese imperialism is extremely decayed, having become a complete obstacle to the development of productive forces.
I have seen a similar report in The New York Times, but when they asked DS about some mass movements in China, DS, during deep contemplation, could mention the White Paper Movement, the 2022 Urumqi疫情, and even the 64事件, but in the end, when answering, they would avoid and say they couldn’t answer. For some sensitive political issues, during deep thinking, DS would even be thinking “How can I express XX in a non-sensitive way,” and then continuously rephrase the wording to be more euphemistic and abstract.
Deepseek often takes too long to load when answering questions, and finally prompts that the server is busy. The solution is often to open another conversation, which is very unfavorable for problems that require continuous questioning to solve. I used to think it was because the server of this thing was too garbage, but now it seems that it might be because it needs to pay extra attention to sensitive issues related to the government, leading to prolonged thinking times.
From a technical perspective, DeepSeek’s review should be comprehensive in many aspects, sometimes directly retrieving sensitive words and responding with “no comment” or a “standard answer,” and sometimes having some specifics during the thought process but then stopping the thinking afterward.