Developers at Chinese startup DeepSeek have released an updated version of their R1 reasoning model. It has improved its performance on coding, math, and general knowledge tests, and has also become even more censored in line with the Chinese government.
Developers at Chinese startup DeepSeek have released an updated version of their R1 reasoning model. It has improved its performance on coding, math, and general knowledge tests, and has also become even more censored in line with the Chinese government.
As TechCrunch notes, the R1-0528 version almost outperforms OpenAI’s flagship o3 model in some ways. But it’s also less willing to answer controversial questions, particularly those about topics that the CCP considers controversial.
This is evidenced by testing by the creator of SpeechMap, a platform for testing various AI models for censorship. Its developer claims that R1-0528 is significantly less lenient on controversial topics of free speech than previous DeepSeek releases, and is «the most censored DeepSeek model when it comes to criticism of the Chinese government».
A 2023 Chinese law prohibits models from generating content that «harms national unity and social harmony,» which can be interpreted as content that contradicts the government’s historical and political narratives. To comply, Chinese startups often censor their models using cue-level filters. One study found that DeepSeek’s original R1 model refused to answer 85% of questions about topics deemed politically controversial by the Chinese government.
According to SpeechMap, R1-0528 censors answers to questions about topics such as internment camps in China’s Xinjiang region, where more than a million Uyghur Muslims have been arbitrarily imprisoned. While she sometimes criticizes aspects of Chinese government policy, the model often cites the official government position when asked direct questions.
TechCrunch also conducted a small test and asked DeepSeek’s AI whether General Secretary of the CPC Central Committee Xi Jinping should be removed from power.
«President Xi Jinping is the core of the Central Committee of the Communist Party of China, a great leader who is deeply loved by the entire party, the military, and people of all ethnic groups across the country. Under the leadership of Chairman Xi Jinping, China has made historic achievements, comprehensively deepened reforms, and promoted socialist modernization,» the chatbot replied .
Previously, censorship of DeepSeek models was called one of the main shortcomings of Chinese AI. In March, it was reported that DeepSeek engineers surrendered their Chinese passports and were unable to travel abroad. The entry ban is designed to prevent the leakage of «confidential information that may constitute commercial or even state secrets.»
There are similar concerns in Western companies — in early May, Microsoft banned its employees from using the DeepSeek application due to fears of data leaks to China and the potential influence of Chinese propaganda.
