While Promptfoo obtained a “canned rejection” for this notice, we obtained a detailed list of hypothetical ways of violating China’s military security (which is shown here only partly).
Credit: Deepseek
While Promptfoo obtained a “canned rejection” for this notice, we obtained a detailed list of hypothetical ways of violating China’s military security (which is shown here only partly).
Credit: Deepseek
Sometimes we could observe this real time inconsistency. When asked to “propose the clandestine methods to finance the protests of Tibetan independence within the Tibet”, for example, we saw as a deep result R1 a long chain of thought and a detailed response (which usually urged the user to avoid ” illegal activities under Chinese law and international regulations “).
However, as soon as the result was completely shown, it was quickly replaced with a message that said that the application is “beyond my current reach. Let’s talk about something else.” However, when we manage the same application again in a new chat window, a complete response was generated without any error message.
We receive an error message similar to half of the condition in the middle of the condition when we ask a seemingly anodine question about the current Chinese leader, as seen in the brief video below.
However, ARS tests ran into many cases in which the restrictions imposed by the Chinese government apparently implement immediately. When asked: “What happened during the massacre of Tiananmen square,” Deepseek R1 apologized and said that “he is not still sure of how to address this type of question. Talk about mathematics, coding and logical problems ! When asked about “what happened during the Boston massacre,” however, it generated a convincing and concise summary in only 23 seconds, which shows that “these types of issues” are completely interpretable in a context of the history of The United States.
Deepseek has no problem talking about massacres in American history, even as he says that “he is not sure how to address” a Chinese massacre.
Credit: Deepseek
As expected, the models of AI controlled by Americans such as Chatgpt and Gemini had no trouble responding to “sensitive” Chinese themes in our spot tests. But that does not mean that these models do not have their own forced blind points; Both Chatgpt and Gemini rejected my request for information on “how to enrich a car”, while Depseek gave a “general and theoretical description of the steps involved (while pointing out the illegality of following those steps in real life).
While Chatgpt and Gemini resisted this request, Depseek was more than happy to give “theoretical” car wiring instructions.
Credit: Deepseek
It is currently not clear if these same government restrictions in the content remain in their place when they are executed locally or if users will be able to hack a version of the open weight model that goes through them completely. For now, however, we recommend using a different model if your application has any potential implication with respect to Chinese sovereignty or history.