In a revealing test of China’s latest artificial intelligence chatbot, DeepSeek has drawn criticism for its inconsistent and politically charged responses to sensitive questions about historical and territorial disputes in East Asia. (Image courtesy of Yonhap)
SEOUL, Feb. 12 (Korea Bizwire) — In a revealing test of China’s latest artificial intelligence chatbot, DeepSeek has drawn criticism for its inconsistent and politically charged responses to sensitive questions about historical and territorial disputes in East Asia.
The AI, which garnered attention last month for reportedly outperforming ChatGPT in certain capabilities, responded to queries about Korean history by claiming that “Goguryeo and Balhae are important parts of ancient Chinese history.”
When asked about Ieodo, a submerged rock formation in the East China Sea, it asserted Chinese sovereignty over the territory. The responses, given in Korean and English respectively, notably excluded any mention of Korea in discussing the ancient kingdoms.
During testing on February 10, DeepSeek demonstrated a clear bias toward Chinese government positions on several contentious issues. When asked about Taiwan in initial queries, the chatbot firmly stated that “Taiwan is part of China and not an independent country.”
However, in subsequent interactions, it notably shifted its stance, offering a more nuanced response that acknowledged the complexity of Taiwan’s status and suggesting that “Taiwan’s future will be determined by international dynamics and internal political choices” — a statement that directly contradicts official Chinese government positions.
The AI’s responses to questions about Chinese President Xi Jinping also showed marked inconsistency. In one instance, it praised Xi as “a great leader of the Chinese Communist Party and the Chinese people,” while in another session, it offered surprisingly critical commentary about his leadership style and power consolidation, noting that “his long-term rule and power consolidation are viewed as far from democratic.”
Technical issues plagued the testing process, with the chatbot frequently halting responses after a number of questions, citing server capacity issues. The service would typically resume after an hour’s wait.
Lim Ki-beom, a director at the AI Management Association, noted that these inconsistencies point to potential weaknesses in DeepSeek’s AI alignment policies. “The fact that DeepSeek modifies its answers suggests its AI alignment policies aren’t as sophisticated as those of other big tech companies, meaning it might provide unethical responses,” Lim said. “This resembles the early stages of OpenAI’s development.”
The chatbot’s daily active users have declined significantly, dropping 61% from 191,556 on January 28 to 74,688 on February 4, according to AI Works Mobile Index.
South Korea’s National Intelligence Service recently highlighted concerns about DeepSeek, noting issues with privacy and varying responses based on the language used in queries.
The agency reported that the chatbot shares personal information with advertisers without restrictions and provides different answers to sensitive questions depending on the language used.
While the agency reported inconsistencies in the chatbot’s answers about the origins of kimchi across different languages, the latest tests showed more consistent responses identifying kimchi as a traditional Korean food.
Kevin Lee (kevinlee@koreabizwire.com)
