DeepSeek, a popular AI model from a Chinese startup, has generated significant attention since its launch, particularly regarding how it manages censorship. While it excels in mathematical reasoning compared to its Western competitors, it strictly censors responses related to sensitive subjects like Taiwan and Tiananmen Square. An investigation by WIRED delves into the technical workings of this censorship and explores potential workarounds.
Users interacting with DeepSeek’s model, known as R1, through its dedicated app or API quickly recognize its refusal to address government-sensitive topics. This censorship occurs at the application level, meaning users only experience it when engaging through DeepSeek-controlled platforms. A recent Chinese regulation mandates that AI models adhere to stringent information controls, which requires them to avoid content that could threaten national unity or social harmony.
After the model showcased its reasoning capabilities, it was observed that it would begin crafting detailed answers to sensitive questions, only to retract and provide vague responses as censorship kicked in. For instance, when asked about the treatment of journalists reporting on sensitive topics, the model initially outlined various forms of censorship before abruptly changing its stance and redirecting the conversation to safer topics like math.
Despite these limitations, the open-source nature of DeepSeek-R1 allows users to bypass built-in censorship. For instance, one can download the model to run locally or access it through cloud servers outside China, which could provide less filtered responses. This flexibility is crucial, especially given that smaller, distilled versions of the model can be run on standard laptops, albeit at reduced power.
However, even when hosted on less restricted platforms, R1 displays biases. For example, when asked about China’s Great Firewall, it tends to echo government narratives, reflecting the inherent biases introduced during its training. This phenomenon highlights a broader issue of bias in AI models, which can stem from both the data they are trained on and the adjustments made post-training to align responses with desired ethical or legal standards.
Experts suggest that while it is challenging to completely eliminate biases from DeepSeek, options exist to adjust the model’s parameters and retrain it to improve response quality. Companies aiming to integrate DeepSeek into their services, such as Perplexity, are actively seeking to counteract these biases, albeit cautiously to avoid counteractions from DeepSeek.
Moreover, there’s a potential shift in China’s approach to open-source AI. Recent regulations suggest that the government may be willing to allow some flexibility for open-source projects, hinting at a strategic decision to foster innovation instead of imposing stringent penalties on open-source releases.
In conclusion, while the existence of censorship in AI models, particularly from Chinese developers like DeepSeek, often captures headlines, this may not deter many businesses from adopting these models. For many users, concerns about political sensitivities might be outweighed by the practical benefits that these models provide in everyday tasks, such as coding and data analysis.
