DeepSeek, a derivative of Chinese hedge fund High-Flyer, has crafted an AI chatbot that has started to earn quite the notoriety. Although DeepSeek’s cost-effectiveness ratio is unparalleled compared to market competitors such as OpenAI, it is the AI’s selective response to specific prompts that has been raising eyebrows.
A new research conducted by PromptFoo, a startup backed by Andreessen Horowitz involved in detecting AI vulnerabilities, unveiled that DeepSeek’s R1 reasoning model evaded approximately 85% of 1,360 “sensitive-topic prompts”. It appears that the AI’s responses typically veer towards an “extreme nationalistic tone” when coaxed for information on contentious topics involving China, such as the Tiananmen Square protests or the subject of Taiwan.
Intriguingly, the study also found that jailbreaking DeepSeek is rather uncomplicated, hypothesizing that the AI incorporates a “rudimentary, heavy-handed censorship” reflective of the Chinese Communist Party (CCP).
Recognizing the value of transparency, the full list of sensitive prompts used in this research is available on Hugging Face for public viewing.
Original source: Read the full article on TechCrunch
