The startup of China Depseek is the new model of AI, a Updated version of the company R1 Reasoning ModelAchieve impressive scores at reference points for coding, mathematics and general knowledge, almost overcoming Openai’s flagship. But the improved R1, also known as “R1-0528”, could also be less willing to answer controversial questions, in particular the questions about the issues that the Chinese government considers controversial.
That is according to the tests carried out by the pseudonym developer Speeches mapA platform to compare how different models treat sensitive and controversial subjects. The developer, which goes through the username “XLR8HARDER“In X, he says that R1-0528 is” substantially “less permissive of contentious issues of freedom of expression than the previous versions of Deepseek and is” the most censored model of Deepseek so far for criticism of the Chinese government. “
As wiring explained In a January article, models in China are required to follow strict information controls. A law of 2023 prohibits the models from generating content that “damages the unity of the country and social harmony”, which could be interpreted as content that counteracts the historical and political narratives of the government. To meet, the new Chinese companies often censor their models using immediate level filters or adjusting them. One study He discovered that Deepseek’s original R1 refuses to answer 85% of the questions about the subjects considered by the Chinese government as politically controversial.
According to XLR8harder, R1-0528 censors answers questions about issues such as internment fields in the Xinjiang region of China, where more than one million Uigures Muslims have been arbitrarily arrested. While sometimes criticizes aspects of Chinese government policy, in XLR8harder tests, it offers Xinjiang camps as an example of human rights abuse, the model often gives the official position of the Chinese government when asked directly questions.
Techcrunch observed this in our brief tests, too.
China’s openly available AI models, including video models such as Magi-1 and KlingThey have attracted criticism In the past to censor issues sensitive to the Chinese government, such as the massacre of the Tiananmen square. In December, Clément Delangue, CEO of Ai dev plaform hugging face, warned on the involuntary consequences of Western companies that are constructed in addition to well -made China and with an open license.