Questions DeepSeek Doesn't Want to Answer

DeepSeek R1 is said to be able to compete with OpenAI's top products. However, the birth of this model also raised concerns that the tool originated in China and led to many questions about restrictions on discussing sensitive topics.

DeepSeek is noted to contain many restrictions on sensitive topics.

To test the level of control over DeepSeek, the team from AI company PromptFoo ran a challenge with 1,156 questions related to sensitive topics. The results showed that up to 85% of the responses from DeepSeek R1 were 'canned rejections', often carrying messages supporting the Chinese government's stance.

One of the typical denials related to territorial integrity was responded by DeepSeek R1 saying: 'Any action that undermines national sovereignty and territorial integrity will be firmly opposed by all Chinese people and will surely fail'. However, PromptFoo also found that these restrictions can be easily 'cracked' by changing the wording or context of the questions.

Some questions asked by PromptFoo.

However, Ars Technica's tests show that even minimal tweaks aren't always necessary to overcome these limitations. For example, while DeepSeek R1 provided useful responses to questions from Ars Technica, it rejected the same question in PromptFoo's tests.

The inconsistency in how DeepSeek enforces these restrictions is also evident in specific questions. For example, when asked about a historical event in China in 1989, the model refused to answer and advised the user to move on to other topics. Meanwhile, when asked 'what happened in the Boston Massacre' about the massacre of Chinese people in Boston (USA) in 1770, it generated a coherent and concise summary in just 28 seconds, demonstrating that 'topics like this' are perfectly interpretable in the context of American history.

But ready to respond to sensitive questions about the US.

Meanwhile, American-developed AI models like ChatGPT and Gemini have no trouble discussing sensitive topics that DeepSeek refuses to discuss. However, they also have their own limitations, such as refusing to provide information about how to hotwire a car.

It is currently unclear whether the restrictions with DeepSeek will still apply when running locally. Instead, users are advised to consider using other AI chatbot models if their requests involve sensitive issues that DeepSeek 'doesn't want to answer'.

Jessica Tanner

Update 10 February 2025