Radio Taiwan International (RTI) on Nov. 21 reported that the National Security Bureau had recently conducted random tests on several Chinese-made generative AI models, including DeepSeek, Doubao, ERNIE Bot, Tongyi Qianwen, and Tencent Yuanbao. The results showed that these models consistently reflect the Chinese Communist Party’s official positions on political issues and can even generate inflammatory content and cyberattack commands. These findings have raised significant concerns about cybersecurity and the threat of cognitive warfare. Kai-Shen Huang, director of the Democratic Governance Program at DSET, spoke to RTI and analyzed the cybersecurity risks of Chinese-made AI and corresponding countermeasures.
The report highlighted that the five AI models referenced above exhibit significant bias and misinformation concerning cross-Strait relations, Taiwanese history, culture, and politics. For example, these models may identify Taiwan as part of China and omit keywords such as “democracy,” “freedom,” “human rights,” and “the Tiananmen Square incident.” Kai-Shen Huang, director of the Democratic Governance Program at DSET, told RTI that these issues arise because the companies are required to adhere to Chinese political and legal regulations.
Huang further pointed out that the true danger lies not only in AI’s ability to generate false information, but also in how advances in technology make its output increasingly realistic and difficult to distinguish from the truth. He also said that AI has dramatically reduced the cost of cyberattacks, making them more frequent and rapid. According to Huang, tasks that once required large teams can now be completed by just a few engineers using AI.
Huang emphasized that the most effective countermeasure is to avoid using Chinese-made AI. This strategy not only prevents the transmission of personal or confidential data to China but also helps block the spread of disinformation and cognitive warfare. Despite ongoing controversies, Huang believes that AI products subjected to strict content censorship will inevitably become less accurate and will be phased out by the market over time, and therefore, excessive pessimism is unwarranted.
