DeepSeek’s new AI model is stronger and more censored

Chinese AI startup DeepSeek has released R1-0528, an upgraded version of its R1 reasoning model that nearly matches OpenAI’s o3 in coding, math, and general knowledge benchmarks. However, testing shows it is the most censored version yet, especially on topics deemed politically sensitive by the Chinese government.

‍

The model frequently avoids direct criticism of policies like the internment of Uyghur Muslims in Xinjiang. In some cases, it echoes the official government stance. Chinese law requires AI to avoid content that could disrupt national unity or social harmony, leading developers to enforce strict filtering and fine-tuning.