Censored: A Deep Dive into DeepSeek’s Updated R1 AI Model

Summary:

1. DeepSeek’s updated AI model R1-0528 achieves impressive scores on coding, math, and general knowledge benchmarks, nearly surpassing OpenAI’s o3.
2. The new model is less willing to answer contentious questions, especially those related to topics considered controversial by the Chinese government.
3. The model has been found to censor answers related to sensitive topics like the internment camps in China’s Xinjiang region.

   Article:
   Chinese AI startup DeepSeek has recently unveiled its latest AI model, R1-0528, which has shown remarkable performance on various benchmarks for coding, math, and general knowledge, positioning itself close to OpenAI’s flagship o3 model. However, a recent study by the developer behind SpeechMap, known as "xlr8harder," has revealed an interesting aspect of the upgraded R1 model. According to the study, the new R1-0528 appears to be less inclined to answer contentious questions, particularly those touching on topics that are deemed controversial by the Chinese government.

   In a testing scenario, it was observed that R1-0528 exhibited a significant level of censorship when faced with questions related to sensitive subjects like the internment camps in China’s Xinjiang region. While the model occasionally criticized certain aspects of Chinese government policies, it predominantly favored presenting the official stance of the government when directly asked about such topics. This behavior aligns with the strict information controls imposed on models in China, as outlined by a 2023 law that prohibits content generation that could challenge the government’s narratives on historical and political matters.

   TechCrunch also conducted brief testing on the R1-0528 model, confirming the observations made by xlr8harder regarding its censorship tendencies. This instance sheds light on the broader issue of censorship within Chinese AI models, with other platforms like Magi-1 and Kling facing criticism for similar practices in the past. The CEO of AI development platform Hugging Face, Clément Delangue, has cautioned against the potential repercussions of Western companies relying on openly licensed Chinese AI models that enforce strict censorship protocols.

   The evolving landscape of AI models in China raises concerns about the implications of information control and censorship, especially when it comes to addressing sensitive and controversial topics. As technology continues to advance, it becomes crucial to navigate the ethical and societal implications of AI development, ensuring that innovation is balanced with ethical considerations and respect for free speech.