DeepSeek’s Updated R1 AI Model Faces Censorship Concerns, Testing Reveals
DeepSeek, a Chinese AI startup, has released an updated version of its R1 reasoning model, dubbed “R1-0528.” While the model exhibits impressive performance in coding, math, and general knowledge benchmarks, nearing OpenAI’s o3, tests suggest increased censorship, particularly regarding topics sensitive to the Chinese government.
SpeechMap, a platform for comparing AI model responses to sensitive subjects, conducted these tests. The developer, known as “xlr8harder” on X, asserts that R1-0528 is significantly less permissive on contentious free speech topics than previous DeepSeek models. According to xlr8harder, this version is “the most censored DeepSeek model yet for criticism of the Chinese government.”
Wired reported earlier this year that Chinese AI models are subject to stringent information controls. A 2023 law prohibits models from generating content that “damages the unity of the country and social harmony,” potentially encompassing anything contradicting official government narratives. Chinese companies often censor their models through prompt-level filters or fine-tuning.
A previous study indicated that the original DeepSeek R1 refused to answer 85% of questions on politically controversial topics as defined by the Chinese government.
Xlr8harder’s testing revealed that R1-0528 censors responses to questions about topics such as the internment camps in China’s Xinjiang region, where over a million Uyghur Muslims have been detained. While the model may sometimes offer criticisms of Chinese government policy, it often defaults to the government’s official stance when directly questioned.
TechCrunch’s brief testing corroborated these findings.
Concerns about censorship in Chinese AI models are not new. Video-generating models like Magi-1 and Kling have previously faced criticism for censoring topics sensitive to the Chinese government, such as the Tiananmen Square massacre. Clément Delangue, CEO of Hugging Face, has cautioned against Western companies building on top of potentially censored Chinese AI models.
