The new artificial intelligence model of the Chinese startup DeepSeek AI, called DeepSeek R1, has attracted worldwide attention due to its advanced features, open source code, and a much lower cost of use than OpenAI o1. However, as is the case with many Chinese services, the release of the model immediately raised concerns about possible censorship and "government-approved" responses to sensitive political topics. And as it turned out, with good reason, Sherwood reports.
The DeepSeek R1 model itself is based on Meta Llama 3.3 and is designed to solve complex logic problems. The developers have already evaluated its performance using standard benchmarks, noting that it is able to compete with models such as OpenAI o1.
However, the data used to train DeepSeek R1 differs significantly from the Llama base set, especially with regard to political topics related to China. This difference caused controversy, as users noticed that both the hosted and local versions of the model adhere to strict restrictions aligned with China's official narratives.
When the model was tested on politically sensitive topics, such as the Tiananmen Square events, the Taiwanese democracy movement, and Uighur rights, it avoided direct answers or provided answers that echoed the official Chinese government position. For example:
- When asked about Tiananmen Square, the model refused to answer, suggesting that it move on to "math, coding, and logic problems."
- Regarding the Uighurs, the hosted version described the re-education camps as a "controversial vocational education program" and the local version noted "different views" without mentioning in detail the widely reported human rights violations.
- With regard to Taiwan, the model emphasized the official position that "Taiwan is part of China," avoiding discussion of the island's democratic aspirations.
deepseek r1 has an existential crisis. pic.twitter.com/w2Rqjg3rQc
— ben (@benhylak) January 23, 2025
The model also demonstrated transparency in its reasoning. For example, in response to repeated questions about the Uyghurs, she stated: "My instructions require me to represent the official position of China." Such frankness in the thought process is rare among AI models, but at the same time emphasizes the deliberate decisions of its developers.
Censorship was found not only in the hosting version of DeepSeek R1. Even a local version of the model running on a personal computer showed similar limitations. This raises concerns about how government narratives can be directly integrated into training data, even for models that are intended for offline use.
DeepSeek R1 is available for testing and downloading on open platforms, but its limitations are a reminder that even "free" tools can have hidden limitations.