Users have reported that DeepSeek refuses to respond to prompts involving politically sensitive topics, such as the Tiananmen Square massacre or critiques of President Xi Jinping. This behavior indicates that censorship is embedded not only at the application level but also within the model’s training data.
A study found that DeepSeek’s R1 reasoning model declined to answer approximately 85% of prompts related to sensitive subjects, often responding with a pronounced nationalistic tone.
Government Regulations Enforce AI Censorship
In April 2023, the Cyberspace Administration of China (CAC) issued draft measures requiring tech companies to ensure that AI-generated content aligns with CCP ideologies, including Core Socialist Values.
These regulations mandate that AI service providers audit generated content and user prompts, either manually or through technical means, to prevent the dissemination of prohibited information.
Such measures compel companies to embed censorship mechanisms within their AI models, ensuring compliance with state directives and reinforcing the CCP’s control over information.
International Concerns and Responses
The integration of censorship into Chinese AI models has raised global concerns about the spread of authoritarian information control. Clement Delangue, CEO of Hugging Face, warned that the widespread adoption of Chinese open-source AI models could inadvertently propagate censorship worldwide.
