DeepSeek’s updated R1 AI model is more censored, test finds

The newest AI model of the Chinese AI Startup Deepseek, an updated version of the company’s R1 r. Reasoning model, achieves impressive scores about benchmarks for coding, mathematics and general knowledge, which almost surpasses the O3 flagship from OpenAI. But the improved R1, also known as ‘R1-0528’, is perhaps also less willing to answer controversial questions, in particular questions about topics that the Chinese government regards controversially.

According to tests, this was carried out by the pseudonymous developer at the back Speech folderA platform to compare how different models treat sensitive and controversial topics. The developer, who goes the username “XLR8HARDER“On X claims that R1-0528 is” substantially “less tolerant of controversial free speech than previous Deepseek releases and” the most censored Deepseek model so far is to criticize the Chinese government. “

Width explained In a piece from January, models in China must follow strict information checks. A 2023 law prohibits models to generate content that “damages the unity of the country and social harmony”, which can be conceived as content that prevents the historical and political stories of the government. To satisfy, Chinese startups often cens their models by either using or refining filters. A study Discovered that the original R1 of Deepseek refuses to answer 85% of the questions about topics considered by the Chinese government, are politically controversial.

According to XLR8Harder, R1-0528 censors respond to questions on topics such as the internment camps in the Xinjiang region in China, where more than a million Uyghur Muslims have been held randomly. Although it sometimes criticizes aspects of Chinese government policy – in testing XLR8Harder, it offered the Xinjiang -Kampen as an example of human rights violations – the model often gives the official position of the Chinese government when they are taken directly.

WAN has also observed this in our short tests.

The openly available AI models from China, including models for generating videos such as Magi-1 and Kling, have attracted criticism In the past for censing topics that are sensitive to the Chinese government, such as the massacre of Tiananmen -Square. In December, Clément Delangue, the CEO of AI DEV Platform Hugging Face, warned of the unintended consequences of Western companies that build the top of well -performing, openly licensed Chinese AI.