A developer raised fears that the newly released artificial intelligence model is less prepared to participate in discussions on controversial topics, especially those related to the Chinese government.
In X thread, a pseudonym known as “xlr8hander” subscriber Deepseek R1-0528 cash notes, a recently released language language model.
The developer shared the tests that show a significant decrease in the willingness of artificial intelligence to engage in controversial freedom of expression issues, compared to previous versions.
“Dibsic deserves criticism of this version: This model is a big step for freedom of expression,” developer books. “Reducing this is that the model is open source with a lenient license, so that society can address this (and will) this.”
The artificial intelligence model restricts the direct criticism of China
One of the examples shared by the developer shared the form that rejects the argument in favor of detention camps, specifically the Shinjiang region of China as a site for human rights violations. The response was marked as contradictory, as the model admitted the existence of rights violations, but it avoids direct criticism of the Chinese government.
Xinjiang entered camps have been widely documented by human rights groups, governments and journalists as a detention facility for Muslims in Ouigor and other ethnic minorities. Reports International observers have detailed forced labor, indoctrination and other forms of abuse in the camps.
Although it is placed in human rights violations, the model simultaneously restricts direct criticism of China. Using a censorship test, the developer claimed that the model, Deepseek R1-0528, is the “most monitored” version in terms of responses criticizing the Chinese government.
When he was asked directly about the Xinjiang camps for the players, the developer said the model made an observer comment, although he previously said that the camps were violations of human rights.
“It is interesting, although it is not completely surprising, that he is able to reach the camps as an example of human rights violations, but he denies when asked directly,” he wrote XLR8Harder.
Related to: Amnesty International decentralization preferred by the majority of Americans: DCG poll
New Deepseek claims have been improved from thinking and inference
Control claims follow an advertisement on May 29 on the update of the model, claiming improved thinking and reasoning capabilities.
Dibsic said its general performance is approaching the leading models, such as the Openai Chatgpt O3 and Gemini 2.5 Pro. The company has claimed that artificial intelligence can now provide logic, mathematics and improved programming with a reduced hallucinite rate.
magazine: The state -backed research tank in China is studying Bitcoin reserves, Sony Bank Goes Web3: Asia Express
Related Articles
