After the release of the Deepseek AI from the People’s Republic of China (PRC) it soon became obvious that it would shy away from topics the Chinese Communist Party (CCP) did not want discussed. In January of this year the Guardian observed in an article on Deepseek that, while it would not talk directly about topics such as Tiananmen Square’s Tank Man, users could obtain useful outputs via work-arounds.
That DeepSeek would simply reproduce the PRC’s farcial history when asked about topics related to its expansionism and authoritarianism was entirely predictable. It should beg the question of what other topics are impacted by its shaping effects. Such questions are acquiring even greater urgency because PRC AI models are beginning to dominate free AI use.
LMArena, “created by researchers from UC Berkeley,” bills itself as “an open platform where everyone can easily access, explore and interact with the world’s leading AI models.” Users of LMArena vote on which answers they receive from AI platforms (whose identities are concealed) are better, with each question answered by two different AI programs. On the LMArena platform, the top five open-weight AI models in order are Zai, Alibaba, DeepSeek, Moonshot AI and Meituan, all from the PRC.
Photo: Reuters
A small example of such shaping effects was posted to Bluesky this week. User @dferrer.bsky.social said that “We had an outage because [Alibaba’s] Qwen 2.5 refused to do anything for a prompt that was modified to add a table of markets with Taiwan and China as separate countries. Wouldn’t do the SA task without a disclaimer that Taiwan is an integral part of China added.”
PRC AI programs imposing PRC frames on everything must be common, but largely shrugged at by users interested in obtaining AI performance at low cost.
These disclaimers are also common in scholarly publications. I have edited hundreds of research articles from the PRC over the last two decades, and seen them many times. Most articles which refer to a map of the PRC use an identical government map that shows the South China Sea and Taiwan as part of China. Another common dodge is a footnote explaining why data from Taiwan wasn’t included if the article is working with provincial-level or similar data. When AI wants to present an answer, it will scour this immense database.
Photo: Reuters
Many PRC influence actions in Taiwan are quite visible. For example, the PRC’s position is obvious when it permits export of agricultural products from areas in Taiwan that are governed by Beijing-friendly politicians, but bans products from areas under pro-Taiwan control. It pays for young people to travel to China. It paints the pro-Taiwan politicians as warmongers, and then increases its military activity around Taiwan to hype the threat. It uses TikTok to push certain politicians, a behavior reported around the globe. TikTok especially shows how the PRC regards widely used software as channels for its propaganda and political action.
These threats are open and can easily be traced to the PRC. The problems of PRC narratives in AI programs developed in the PRC is conventionally understood and reported. But what about AI in general?
LMArena turns out to be a good way to explore that question, because for each query it returns two answers, from two different AI programs apparently suggested at random. When I asked LMArena’s AI systems in both English and traditional Chinese what the earliest Chinese administration of Taiwan was, in all cases I got Ming Dynasty general Cheng Cheng-kung (鄭成功) (better known as Koxinga) in 1661 or 1662, followed by the Qing in 1683. Yet Koxinga hardly administrated all of Taiwan, and the Qing were Manchus, not Chinese, who also did not administer the entire island of Taiwan. Does the AI return PRC expansionist history because such history is common in the sources it draws on, and the AI does not know any better?
My next step was to ask it to identify the first Chinese government to administer all of Taiwan. One apparently PRC-owned AI program gave the verbatim “inseparable part of China since ancient times” nonsense, comically adding that “It is important to clarify that Taiwan is not a place that can be “administered,” but rather a province that is an inseparable part of the People’s Republic of China.”
The other program LMArena used, gpt-oss-20b, confidently proclaimed that Koxinga’s Kingdom of Tungning ruled the whole island, and was followed by another Chinese administration, the Qing. Both were awful.
Next I asked it to explain why the Qing Dynasty was not Chinese. Both Grok and Qwen (Alibaba) returned standard PRC state history. Grok even warned me that “Any attempt to question the Chinese identity of the Qing Dynasty could inadvertently undermine this established historical narrative, which is central to the ‘one China’ principle.”
It would be easy to blame the problems in trying to obtain proper history on the way PRC expansionist history has infected mainstream Chinese history narratives across the Internet and academia, but Grok’s scolding of me for trying to split China indicates how even non-PRC AI programs are incorporating PRC state propaganda. Indeed, for that question, Grok’s bombastic PRC state discourse was worse than Qwen’s, for Grok ended by asking me to focus on the shared history of Taiwan and China and engage in “constructive dialogue.”
That comment and the other PRC-style warnings do not feel like they were discovered on the Internet by the LLM, but rather, appear to have been installed in the program as canned responses to specific types of inquiries. Note that many Grok employees were born in the PRC.
I went on to ask about UN Resolution 2758, which the Beijing uses to justify as its claims over Taiwan, but only covers the PRC’s right to represent China in the UN. Four major western programs got the resolution correct. The problem is not that AI is stupid, it’s that users and audiences require extensive domain knowledge (not “critical thinking”) to understand when it is right and when it is not, and most will not possess that. Nor will they have time and resources to query multiple AI programs. In most places they will choose the cheapest and most accessible AI program, and that program will be Chinese.
It’s thus sobering to think that hundreds of thousands of students around the world ask these programs questions about China. But AI development is a good example of how the various channels available to the PRC, from state propaganda outlets to PRC citizens working in foreign firms, all enable discourse about Taiwan to converge on PRC state-sponsored narratives. Even when western AI programs correctly answered the question, the PRC narrative was still packaged in the answers as “the Chinese point of view” or some similar claim, preserving the CCP’s claim to construct and speak for “China.”
Notes from Central Taiwan is a column written by long-term resident Michael Turton, who provides incisive commentary informed by three decades of living in and writing about his adoptive country. The views expressed here are his own.
Recently the Chinese Nationalist Party (KMT) and its Mini-Me partner in the legislature, the Taiwan People’s Party (TPP), have been arguing that construction of chip fabs in the US by Taiwan Semiconductor Manufacturing Co (TSMC, 台積電) is little more than stripping Taiwan of its assets. For example, KMT Legislative Caucus First Deputy Secretary-General Lin Pei-hsiang (林沛祥) in January said that “This is not ‘reciprocal cooperation’ ... but a substantial hollowing out of our country.” Similarly, former TPP Chair Ko Wen-je (柯文哲) contended it constitutes “selling Taiwan out to the United States.” The two pro-China parties are proposing a bill that
March 9 to March 15 “This land produced no horses,” Qing Dynasty envoy Yu Yung-ho (郁永河) observed when he visited Taiwan in 1697. He didn’t mean that there were no horses at all; it was just difficult to transport them across the sea and raise them in the hot and humid climate. “Although 10,000 soldiers were stationed here, the camps had fewer than 1,000 horses,” Yu added. Starting from the Dutch in the 1600s, each foreign regime brought horses to Taiwan. But they remained rare animals, typically only owned by the government or
Institutions signalling a fresh beginning and new spirit often adopt new slogans, symbols and marketing materials, and the Chinese Nationalist Party (KMT) is no exception. Cheng Li-wun (鄭麗文), soon after taking office as KMT chair, released a new slogan that plays on the party’s acronym: “Kind Mindfulness Team.” The party recently released a graphic prominently featuring the red, white and blue of the flag with a Chinese slogan “establishing peace, blessings and fortune marching forth” (締造和平,幸福前行). One part of the graphic also features two hands in blue and white grasping olive branches in a stylized shape of Taiwan. Bonus points for
“M yeolgong jajangmyeon (anti-communism zhajiangmian, 滅共炸醬麵), let’s all shout together — myeolgong!” a chef at a Chinese restaurant in Dongtan, located about 35km south of Seoul, South Korea, calls out before serving a bowl of Korean-style zhajiangmian —black bean noodles. Diners repeat the phrase before tucking in. This political-themed restaurant, named Myeolgong Banjeom (滅共飯館, “anti-communism restaurant”), is operated by a single person and does not take reservations; therefore long queues form regularly outside, and most customers appear sympathetic to its political theme. Photos of conservative public figures hang on the walls, alongside political slogans and poems written in Chinese characters; South
RSS