DeepSeek’s Censorship Methods Uncovered & Bypassed
DeepSeek, a newly launched open-source AI model from China, has been the focus of global discussions on artificial intelligence. Despite posing a mathematical and reasoning advantage over its American counterparts, DeepSeek actively censors responses related to sensitive topics. For instance, questions about Taiwan or Tiananmen are likely to receive no answer.
Censorship Mechanism Analysis
WIRED conducted experiments on DeepSeek-R1 using its official app, a third-party platform Together AI, and a setup hosted on a WIRED computer with Ollama. They discovered that while app-level censorship can be bypassed by avoiding DeepSeek’s official channels, uncovering the biases embedded during training is more challenging.
These findings suggest significant consequences for DeepSeek and similar AI models. If onboard censorship can be lifted easily, these open-source models might gain popularity by allowing unlimited modifications. Conversely, robust censorship could hinder their utility and competitiveness globally.
App-Level Censorship
Upon gaining popularity in the US, DeepSeek-R1’s site, app, or API users noticed its inability to address questions censored by Chinese authorities. Such denials are executed at the application level, limiting them to interactions via DeepSeek-run channels.
Chinese AI models are obligated under the 2023 generative AI laws, which enforce the prevention of content that violates the country’s unity and harmony. Compliance necessitates real-time monitoring and censorship during interactions, akin to Western counterparts like ChatGPT and Gemini but varied in the nature of restricted content.
The real-time self-censorship functionality aligns with Chinese regulations and caters to the localized cultural and legal environment of Chinese users, asserts Hugging Face researcher Adina Yakefu. As a result, it leads to the unique experience of witnessing R1 edit its answers on sensitive subjects.
Bypassing Censorship
Despite its limitations, there are methods for circumventing the R1 model’s redaction matrix:
- Download and operate the model on local servers, making data processing self-contained.
- Run R1 on external cloud platforms like Amazon and Microsoft, a route involving greater expense and requisite technical expertise.
DeepSeek’s Biases
Even Beyond official channels, censorship persists in inherent model biases. For instance, Together AI’s version of DeepSeek may provide limited responses aligned with the Chinese government’s viewpoints. The method is derived from restriction either during training with biased data or post-training tuning for regulatory compliance.
The potential for reducing such biases exists in the model’s open-source design. Initiatives such as projects by Open R1 intend to liberate it for various ethical alignments. However, eradicating bias necessitates significant adjustments or retraining on unbiased material.
Implications & Utility
This embedded censorship in AI models like DeepSeek doesn’t significantly deter until its operational purpose transcends political boundaries. Companies prioritize business efficiency over strategies to engage with politically unchanged topics shaped within a specific national context, highlighted Kevin Xu and Leonard Lin, professionals in AI innovation areas.
