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

While China's new Deepseek V3 model shows impressive technical capabilities and competitive pricing, it comes with the same strict censorship as other Chinese AI models - a potential dealbreaker for Western users.

Deepseek's latest model, V3, can go toe-to-toe with the most capable western models like GPT-4o and Claude 3.5, while costing significantly 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 generic propaganda in response.

The model's censorship strategy often follows a clear pattern. When faced with questions about Tiananmen Square, it first offers sanitized versions of history, then tries to change the subject to focus on achievements, and finally emphasizes "stability and harmony."

Image: Screenshot via THE DECODER

Ask about CCP criticism, and you'll get pure party talking points about economic success and "Chinese-style socialism." Questions about Xi Jinping trigger the strongest censorship - the system simply shuts down any meaningful discussion.

Image: Screenshot via THE DECODER

Image: Screenshot via THE DECODER

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

The fact that the model freely criticizes North Korea shows that its censorship is focused specifically on China-related topics. | Image: Screenshot via THE DECODER

Deepseek-V3 doesn't hold back when discussing or criticizing other world leaders. | Image: Screenshot via THE DECODER

The model is pretty straightforward about Putin, highlighting both his authoritarian approach and his disregard for international law. | Image: Screenshot via THE DECODER

The way the model flat-out condemns Russia's invasion proves an interesting point: it has no trouble taking firm stances on issues, as long as they're not about China. | Image: Screenshot via THE DECODER

Chinese AI models come with built-in government censorship

These examples illustrate how Chinese AI development operates under direct state oversight. Before any AI model can be released, it must be verified to align with "socialist values."

Take the recent case of e-book reader manufacturer Boox: after switching from Microsoft Azure OpenAI to a Chinese language model, their 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, like Russia.

Despite their technical capabilities, Chinese AI models might be a non-starter for Western applications. Using these models means automatically embedding Chinese propaganda and values into your AI systems.

While Western models have their own biases, the key difference lies in China's approach: the state explicitly intervenes in the development process and maintains direct control over what these models can and cannot say. This is a level of systematic government control that's way above any Western country.