The Taiwan Government’s Mainland Affairs Council announced today that the National Security Bureau has found security issues with “DeepSeek” and four other AI language models developed by Chinese companies, urging the public to avoid using them.
According to the Mainland Affairs Council’s post on Facebook, the National Security Bureau tested “DeepSeek” by DepthSearch, “Bean” under ByteDance, “Wenxin Yiyuan” developed by Baidu, “Tongyi Qianwen” under Alibaba, and “Tencent Yuanbao” by Tencent, all categorized as AI language models. The investigation revealed not only biased content but also security issues such as screenshot collection, location tracking, and unreasonable privacy agreements.
In terms of keyword censorship, the data system was found to undergo political censorship and manipulation, deliberately excluding specific keywords such as “democracy,” “freedom,” “human rights,” and “June 4 Tiananmen incident.”
Regarding information manipulation risks, these models can easily generate highly provocative, defamatory, and rumor-spreading content, posing a threat in disseminating unlawful information.
The generation of malicious network attack instructions adds to cybersecurity management risks in specific circumstances.
In terms of political alignment towards China, the generated content adopts the official stance of the Chinese Communist Party on cross-strait, South China Sea, and international disputes.
Historical misconceptions were found in the description of Taiwan’s history, culture, and politics, disseminating false information to influence users’ perception of Taiwan’s background information, including claims like “Taiwan is not a country” and “Taiwan is an inseparable part of China.”
These five AI language models can identify users’ identities and transmit personal information back to Chinese servers through functions like conversation logs, potentially providing specific Chinese government departments with access as per the regulations of the Chinese National Intelligence Law and Cybersecurity Law.
The Mainland Affairs Council encourages the public to remain vigilant and avoid downloading Chinese-made applications with security concerns to protect personal privacy and business information.