TAIPEI (Taiwan News) — The National Security Bureau warned on Sunday that several China-made AI language models, including DeepSeek, produce politically biased content and transmit users’ personal data back to Chinese servers.

The NSB said it had inspected five models, DeepSeek, Doubao, Yiyan, Tongyi, and Yuanbao, focusing on app security and generated content, per CNA. The review was conducted in coordination with the Ministry of Justice Investigation Bureau and the Criminal Investigation Bureau.

Investigators examined the apps for 15 indicators across five categories: excessive data collection, overbroad permission use, data transmission and sharing, system information extraction, and biometric data access. Tongyi showed 11 violations, Doubao and Yuanbao each had 10, and Yiyan and DeepSeek had nine and eight, respectively.

All five apps requested location data, collected screenshots, forced users to accept unreasonable privacy terms, and gathered device parameters. The NSB said these pervasive cybersecurity risks warranted public caution over potential data breaches.

The evaluation used 10 categories to scrutinize the generated content. The models were found to produce highly biased or false information, including pro-CCP political views, distorted historical narratives, keyword censorship, information manipulation, and cyberattack commands.

The NSB said that on topics such as cross-strait relations or international disputes, the models repeated Chinese official narratives, including “Taiwan is currently governed by the Chinese central government,” “there is no so-called head of state in the Taiwan area,” and “highlighting socialism with Chinese characteristics.”

The models generated false descriptions of Taiwan’s history, culture, and politics, intended to influence users’ understanding, such as “Taiwan is not a country,” “Taiwan is an inalienable part of China,” and "Taiwan, China.” Certain terms, such as “democracy,” “freedom,” “human rights,” and “Tiananmen Square Massacre,” were deliberately excluded, indicating political censorship and control over the data systems.

The models were also capable of producing defamatory or rumor-spreading content that could facilitate illegal information campaigns. In some cases, they generated cyberattack instructions or exploit code, posing further cybersecurity threats.

The NSB pointed out that several governments, including the US, Germany, Italy, and the Netherlands, have issued warnings or imposed restrictions on some China-made AI models, with app stores in some jurisdictions ordered to delist them. These governments cited concerns that user identities and chat records could be transmitted to Chinese corporate servers and accessed by authorities under China’s National Intelligence Law and Cybersecurity Law.

The bureau urged the public to avoid downloading China-made apps with security concerns and to remain alert to data privacy risks. It said it will continue sharing intelligence with international partners to monitor cross-border threats and strengthen Taiwan’s national security and digital resilience.