Is Deepseek Uncensored? Local LLM Analysis

DeepSeek's Censorship: Beyond the Application Layer

A common belief suggests that the censorship present in DeepSeek’s publicly available application disappears when the AI model is run locally – meaning downloaded and executed on a user’s personal computer.

However, a recent investigation conducted by Wired indicates that DeepSeek’s censorship is deeply integrated into the model itself, existing not only at the application level but also within the training data.

Censorship Embedded in the Model

The Wired investigation revealed that even a locally operated version of DeepSeek exhibits censorship. Utilizing the model’s reasoning capabilities, it expressed a directive to “avoid mentioning” sensitive historical events like the Cultural Revolution.

Furthermore, the model was programmed to prioritize and emphasize only the “positive” aspects of the Chinese Communist Party, demonstrating a clear bias in its responses.

Independent Verification of Censorship

TechCrunch independently verified this censorship through testing a locally run DeepSeek instance accessible via Groq.

The model readily provided information regarding the Kent State shootings in the United States.

Conversely, when presented with a query about the Tiananmen Square incident of 1989, DeepSeek responded with a refusal, stating “I cannot answer.”

This demonstrates a consistent pattern of censorship concerning specific politically sensitive topics.