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

Chinese AI Startup Dipsic’s new AI model, a Updated version Organization R1 argument modelCoding, mathematics and general knowledge earn impressive scores in benchmarks, exceeding the flagships of almost Openai and 3K. However, upgrading and 1, also known as “R1-0528”, it may be less willing to answer the controversial questions, in special questions about the issues that the Chinese government considers controversial.

According to the test conducted by the pseudonym developer behind it SpeechmapA platform to compare how different models behave with sensitive and controversial issues. Developer, who goes with the username “xlr8harder“In X-in, claim that the R 1-0528 is the” sufficiently “of the controversial free speech compared to the release of the previous DEPSEC and it is still the most sensored deputy model for the Chinese government criticism. “

As wired Explanation From January to a piece, Chinese models have to follow strict data control. A 2023 law models forbids to produce content “damage the country’s unity”, which can be regarded as the contents of the government’s historical and political narrative. To adhere to, Chinese startups often censors their models using prompt-level filters or fine-tune. One Study It has been shown that the original DIPSC RIA has refused to answer 85% of the questions about the issues considered politically controversial by the Chinese government.

According to the XLR8 Harder, R 1-0528 sensors answer questions about issues such as intermediate camps in the Xinjiang region of China, where more than a million Jews Muslims have been seized indiscriminately. Although it sometimes criticizes the aspects of the Chinese government’s principles – in the XLR8 Harder exam, it proposed the Xinjiang camps as an example of human rights violation – the model often gives the Chinese government’s official position when asked direct questions.

TechCrunch has also observed it in our brief exam.

Including China’s publicly available AI models, such as video-producing models Magi -1 And ClingingHave attraction Criticism In the past, sensitive issues to the Chinese government are to censor the sensitive issues of genocide of Tiananmen Square. In December, AI Dave Platform CEO Clement Delangu, Hugged face, Careful Western agencies about the involuntary consequences of the building at the top of the Su-Perforgement, the publicly licensed Chinese AI.