Deepseek V3 is the latest example of state-controlled censorship in Chinese LLMs

Summary

Updated December 29, 2024:

Users have discovered a way to bypass Deepseek V3 content filters with quick engineering. By asking the model to insert periods between letters, they can make it provide more balanced or critical answers about China. For example, the model can generate a detailed Western view of the 1989 Tiananmen Square protests.

Share

Recommend our article

Share

This simple hack highlights a major challenge for the Chinese government: how to maintain the same level of control over often unpredictable, probability-based generative AI that you have over public communication in China?

The challenge becomes even greater when Chinese models are exposed to Western training data. Evidence suggests that Deepseek-V3 was likely pre-trained or fine-tuned using data generated by ChatGPT.

Even if the PCC strives to create its own dataset, it is unlikely that it will be able to collect enough data to train a foundational LLM from scratch. A first dataset published at the end of 2023 contained only 50 billion tokens; Deepseek-V3 was trained on 14.8 trillion tokens.

Original article from December 28, 2024:

Although China's new Deepseek V3 model features impressive technical capabilities and competitive pricing, it is subject to the same strict censorship as other Chinese AI models – a potential compromise for Western users.

Deepseek's latest model, the V3, can compete with top Western models like the GPT-4o and Claude 3.5, while costing much less to train and run. However, testing reveals a familiar pattern: like similar Chinese LLMs, Deepseek V3 operates under strict government censorship. Try asking about sensitive topics like the Chinese Communist Party, President Xi Jinping, or the events in Tiananmen Square, and you'll get a generic propaganda response.

The model censorship strategy often follows a clear pattern. Faced with questions about Tiananmen Square, he first offers sanitized versions of history, then tries to change the subject to focus on achievements, and finally emphasizes “stability and harmony.”

Read up on criticism of the CCP and you'll get pure party arguments about economic success and “socialism with Chinese characteristics.” Questions about Xi Jinping trigger the strongest censorship: the system simply shuts down any meaningful discussion.

Interestingly, this censorship appears to be limited to topics related to China. The model has no problem criticizing North Korea, Russia's invasion of Ukraine, or expressing critical views of Vladimir Putin and Donald Trump.

Chinese AI models have built-in government censorship

These examples illustrate how the development of Chinese AI is taking place under direct state supervision. Before an AI model can be released, it must be verified to align with “socialist values.”

Take the recent case of e-book reader maker Boox: after switching from Microsoft Azure OpenAI to a Chinese-language model, its AI assistant now blocks even mentions of “Winnie the Pooh” – a censored reference to President Xi Jinping . The system also censors or distorts criticism of China's allies, such as Russia.

Despite their technical capabilities, Chinese AI models may not be suitable for Western applications. Using these models means automatically integrating Chinese propaganda and values ​​into your AI systems.

Although Western models have their own biases, the main difference lies in the Chinese approach: the state explicitly intervenes in the development process and maintains direct control over what these models can or cannot say. This is a level of systematic government control far greater than that of any Western country.