Understanding and Circumventing DeepSeek Censorship
Just two weeks after releasing its open-source AI model, DeepSeek remains at the center of discussions about the future of artificial intelligence. Despite demonstrating superiority in mathematics and reasoning over some U.S. competitors, the Chinese startup strategically censors its interactions. Inquiries regarding Taiwan or Tiananmen are often met with silence from DeepSeek R1.
Exploring Censorship Amongst Versions
WIRED conducted an investigation to demystify this censorship by examining DeepSeek-R1 through its own app, a third-party application Together AI, and using Ollama on a WIRED computer. It emerged that sidestepping the straightforward censorship employed on their app is possible by utilizing alternative methods. However, ingrained biases from the training phase present more complex challenges to overcome.
These findings could significantly impact not only DeepSeek but other Chinese AI enterprises as well. Easy dismantling of censorship mechanisms might increase the adoption of Chinese AI models globally. Conversely, stringent filters could decrease competitiveness in the global AI landscape. DeepSeek refrained from commenting to WIRED’s inquiries.
Application-Level Censorship
Upon reaching popularity in the U.S., DeepSeek-R1 consistently started to refuse to address politically sensitive topics like those deemed sensitive by the Chinese government. These refusals are triggered at the application level and exclusively manifest when interacting through channels under DeepSeek’s control.
This practice is prevalent among Chinese language models, especially under the 2023 regulation which mandates adherence to government guidelines applicable to social media and search engines in China. The law restricts AI from creating content jeopardizing national unity and harmony.
Real-Time Content Monitoring
Chinese AI models often monitor and censor interactions live to comply with these regulations. For example, when WIRED asked R1 about the treatment of journalists by authorities, it initially crafted a detailed answer mentioning censorship and detention. Yet, it spontaneously halts and responds with, “Sorry, I’m not sure how to approach this type of question yet. Let’s chat about math, coding, and logic problems instead!”
Circumventing Censorship
R1’s open-source nature, however, presents viable methods to bypass censorship. Users have the option to download and run the model locally, which, despite requiring significant GPU capacity for optimal performance, remains accessible through smaller versions on personal laptops. Alternatively, using cloud services outside China can unlock powerful model versions, albeit costlier and technologically demanding.
Built-In Bias
Together AI’s version, although not outright refusing answers, reveals clear biases prevalent in biased training data, often crowning Chinese governmental views. Responses tend to adhere closely to set narratives by the Chinese state, emphasizing only positive aspects.
Overcoming Pre-Training and Post-Training Biases
DeepSeek’s foundational model can be adapted by removing biases post-training but it’s not an easy task. Techniques proposed by AI scientist Eric Hartford involve adjusting model weights or re-training using a database of censored subjects. Such initiatives benefit significantly from models slightly adjusted on the post-training level, having less inherent bias.
Industry Implications
The open-source model at DeepSeek could potentially become unshackled from all censorship, stirring concerns within Chinese boundaries. However, the governmental thrust seems lenient, perhaps evidenced in DeepSeek’s flourishing despite tight regulations.
Despite noticeable limitations within some contexts—DeepSeek offers considerable practical advantages, prompting expectations of heightened international collaboration devoid of local censorship constraints. Many businesses globally may prioritize operational usefulness over regional political alignments.
