DeepSeek, a Chinese AI startup, has released an updated AI model known as R1-0528, building on the company’s previous R1 reasoning model. This new iteration has achieved notable scores in coding, math, and general knowledge benchmarks, nearly matching OpenAI’s flagship model, o3. However, R1-0528 may be more restrained in addressing contentious topics, especially those considered controversial by the Chinese government.
This observation comes from testing conducted by a pseudonymous developer from the platform SpeechMap, which evaluates how different models handle sensitive subjects. The developer, using the username “xlr8harder” on X, asserts that R1-0528 is considerably less permissive regarding free speech issues compared to earlier DeepSeek models, and it is notably more restricted concerning criticism of the Chinese government.
Wired previously highlighted that models in China must adhere to strict information regulations. A law enacted in 2023 prohibits these models from creating content that could undermine the country’s unity and social harmony. To comply, Chinese companies often censor models through prompt-level filters or by fine-tuning. Research indicated that DeepSeek’s original R1 model refused to address 85% of politically sensitive questions as identified by the Chinese authorities.
According to “xlr8harder,” R1-0528 avoids providing answers about topics like the internment camps in the Xinjiang region, where over a million Uyghurs have been detained. Although the model occasionally acknowledges certain human rights abuses, it often aligns with the Chinese government’s official narratives when directly questioned.
TechCrunch also noted similar findings during its brief tests.
Chinese AI models, including video-generating ones like Magi-1 and Kling, have faced criticism for censoring politically sensitive content. In December, Clément Delangue, CEO of the AI development platform Hugging Face, expressed concerns about the implications of Western companies utilizing Chinese AI models with open licenses due to potential unintended consequences.
