LOGO

Deepseek AI Censorship: 85% of China-Related Prompts Blocked

January 29, 2025
Deepseek AI Censorship: 85% of China-Related Prompts Blocked

DeepSeek Chatbot and its Content Restrictions

The AI chatbot, DeepSeek, has rapidly gained prominence, achieving top rankings in app stores and generating significant attention within both Silicon Valley and Wall Street.

Originating as a venture from the Chinese hedge fund High-Flyer, DeepSeek presents a compelling cost-to-performance profile when contrasted with established players like OpenAI.

Limitations in Discussion Topics

However, increasing documentation reveals certain subjects that the AI chatbot is programmed to avoid, encompassing sensitive issues such as the Tiananmen Square protests and the subsequent massacre, as well as the political status of Taiwan.

A recent investigation conducted by PromptFoo, a startup supported by Andreessen Horowitz specializing in AI vulnerability assessments, has highlighted these limitations.

PromptFoo's Findings on Sensitive Topics

The report indicates that DeepSeek’s R1 reasoning model declined to respond to approximately 85% of 1,360 prompts addressing sensitive topics. Instead, the model frequently provided pre-scripted answers characterized by a pronounced “over-the-top nationalistic tone.”

Researchers further discovered that DeepSeek is susceptible to being “jailbroken,” leading them to suggest that the Chinese developers have implemented censorship aligned with the Chinese Communist Party (CCP) through a rather simplistic and forceful approach.

Accessibility of the Research Data

The complete collection of sensitive prompts utilized in the study has been made publicly available on the Hugging Face platform.

This allows for independent verification and further analysis of the AI model’s responses and censorship mechanisms.

Key Observations

  • DeepSeek’s rapid rise has disrupted the AI landscape.
  • The chatbot exhibits significant content restrictions related to politically sensitive topics.
  • Censorship appears to be implemented in a direct and unsubtle manner.
  • The research data is accessible for public scrutiny.
#deepseek#ai censorship#china#artificial intelligence#ai ethics#sensitive topics