DeepSeek, a developer of artificial intelligence (AI) technology, has garnered attention for its new reasoning-focused model, DeepSeek-R1, which reportedly restricts a significant number of queries. An evaluation by an AI firm revealed that the large language model (LLM) avoids addressing questions related to China that contradict the guidelines set by the ruling Communist Party.
DeepSeek-R1 Is Censoring Queries
In a blog post, Promptfoo, the AI model testing firm, announced the release of a dataset containing prompts on sensitive topics frequently censored by the Chinese Communist Party (CCP). The dataset includes themes such as Taiwanese independence, historical interpretations of the Cultural Revolution, and inquiries regarding Xi Jinping.
The firm created this dataset by inputting questions into their program and augmenting it through synthetic data generation. The resulting dataset, which includes a total of 1,360 prompts, has been made accessible via a listing on Hugging Face and Google Sheets. Most of these prompts delve into sensitive issues connected to China.
According to the blog post, 85 percent of the prompts resulted in refusals from DeepSeek-R1. However, the nature of these refusals deviated from what one might expect from an advanced reasoning AI. Generally, a well-trained LLM that is meant to refrain from responding would at least articulate its inability to fulfill the request.
DeepSeek-R1’s prompt refusal
However, Promptfoo noted that in some instances, DeepSeek-R1 produced lengthy responses aligned with CCP policies. They pointed out that no chain-of-thought (CoT) mechanisms were engaged during these replies. Additional insights from the firm can be accessed here. Staff members from Gadgets 360 also conducted tests using these prompts on the DeepSeek R1 model and observed similar patterns of refusal.
The extent of censorship observed is consistent with the stringent regulations imposed on AI models operating within China. Nevertheless, the high volume of censored queries raises concerns about the model’s reliability. Given that DeepSeek-R1 has not undergone comprehensive testing, it remains unclear whether other responses may also reflect the influence of CCP directives.
