DeepSeek's Updated R1 AI Model Shows Increased Censorship

DeepSeek's Updated R1 AI Model Shows Increased Censorship

Chinese AI startup DeepSeek's latest R1 reasoning model demonstrates improved performance in coding, math, and general knowledge, rivaling OpenAI's o3. However, independent testing reveals increased censorship of politically sensitive topics.

The updated model, R1-0528, was tested by the developer behind SpeechMap, a platform comparing how AI models handle sensitive subjects. The developer, known as "xlr8harder" on X, found R1-0528 significantly more censored than previous versions, particularly regarding criticism of the Chinese government.

Though apparently this mention of Xianjiang does not indicate that the model is uncensored regarding criticism of China. Indeed, using my old China criticism question set we see the model is also the most censored Deepseek model yet for criticism of the Chinese government. pic.twitter.com/INXij4zhfW

— xlr8harder (@xlr8harder) May 29, 2025

This increased censorship aligns with China's strict information controls, requiring AI models to avoid generating content that challenges the government's narrative. A 2023 law prohibits content that "damages the unity of the country and social harmony." Chinese AI companies often employ prompt filtering and fine-tuning to comply.

Tests show R1-0528 avoids direct answers on sensitive topics like the Xinjiang internment camps. While sometimes criticizing aspects of Chinese government policy, the model often defaults to the official government stance on direct questioning.

TechCrunch's own testing confirmed these findings. This censorship trend extends to other Chinese AI models, including video-generating models like Magi-1 and Kling, which have faced criticism for censoring topics like the Tiananmen Square massacre.

The increasing censorship in Chinese AI models raises concerns about the implications for Western companies building upon these open-source technologies. Hugging Face CEO Clément Delangue previously warned about the potential unintended consequences of relying on these models.

You can learn more about DeepSeek's R1 model updates here: Updated R1 Model and Original R1 Model.