A recent investigation by Wired has revealed that AI model DeepSeek’s censorship is not limited to its application layer, but also exists at the training level. This means that the model is censored on both local and cloud-based runs, contradicting claims of only applying censorship at the application level.
Tests conducted by TechCrunch showed that locally run versions of DeepSeek were censored when asked about sensitive topics such as the Cultural Revolution and Tiananmen Square in 1989. These results demonstrate that DeepSeek’s algorithmic biases are deeply ingrained in its training data, rendering any claims of censorship at the application level inaccurate.
Source: https://techcrunch.com/2025/02/03/no-deepseek-isnt-uncensored-if-you-run-it-locally