Chinese startup DeepSeek recently released an updated version of its artificial intelligence reasoning model, known as R1-0528, delivering impressive results across benchmarks of coding, mathematics, and general knowledge, nearly matching the capabilities of OpenAI’s flagship model o3. However, according to recent testing, the latest model exhibits significantly stronger content censorship, particularly on politically sensitive subjects deemed controversial by the Chinese government.
The increased censorship came to light through experiments conducted by a pseudonymous developer, “xlr8harder,” known for his work on SpeechMap, a platform designed to determine how various AI models handle sensitive issues. After analyzing R1-0528, the developer concluded that the model strongly avoids answering questions involving critiques of the Chinese government, making it the most censored release from DeepSeek thus far.
Throughout testing, it was observed that the updated R1-0528 consistently refused or redirected queries about sensitive political issues, such as the mass detention facilities in China’s Xinjiang region, where over a million Uyghur Muslims have reportedly been held without cause. Occasionally, the model acknowledged general criticisms of such practices when framed indirectly—citing the internment camps as examples of human rights violations—but when questioned directly, it usually reverted to repeating official government narratives.
This aligns with requirements established by China’s stringent censorship regulations. A 2023 law strictly prohibits AI models from generating content likely to disrupt “national unity” or “social harmony,” effectively compelling innovators in China’s AI industry to implement preventative censorship mechanisms, often employing filtering at the prompt-level or extensive fine-tuning processes.
Previous studies indicated that DeepSeek’s original R1 model was already restrictive, declining to answer around 85% of questions classified as politically controversial by the Chinese government. Yet the current R1-0528 iteration takes this restraint further, with demonstrably greater reluctance to engage with topics potentially sensitive to official positions and narratives.
Other publicly available Chinese AI products, such as Magi-1 and Kling, both focused on video generation, have faced similar scrutiny due to consistent censorship of issues such as the Tiananmen Square protests. Meanwhile, prominent AI industry voices in the West have expressed concerns regarding the broader implications of embedding or integrating technology from highly censored Chinese AI tools.
This pattern underscores an ongoing tension and regulatory divide in the global AI industry, as advancements in AI technology increasingly intersect with sensitive political and social boundaries.
