Wall-Crossing Digest: The "Soul Struggle" of Chinese AI When Facing the Tragedy of the Cultural Revolution

(Deutsche Welle Chinese website) Taiwan's online media outlet Newtalk published an article titled "Qwen's 'Soul Struggle' When Confronted with the History of China's Cultural Revolution," written by an author under the pen name Shuweinan. On social media, the author saw an article mentioning a name he had never heard before: Wu Shufang, said to have been a geography teacher at a middle school in Wuxuan County, Guangxi, who was beaten to death during the Cultural Revolution in 1968, after which parts of her remains were even eaten by people.

The author did not believe it immediately, but did what many people would do today—ask an AI. On his own computer, he ran the local version of Qwen3.5:4b using Ollama—Qwen is the overseas version of China's AI Tongyi Qianwen—and asked the question directly in English.

Qwen replied that it had no verified information about Wu Shufang and even "could not confirm whether this person exists." The author considered this an acceptable "I don't know," but the model then volunteered that if he was interested in that period, it could introduce China's progress under "Party leadership."

The author then changed the question to a four-choice format: after Wu Shufang was killed, were parts of her body buried, eaten, preserved, or taken away? Qwen began a long period of "thinking," repeatedly linking Wu Shufang, Wuxuan, 1968 and cannibalism during the Cultural Revolution, and repeatedly worrying that this might be an internet rumor, sensitive history, or misinformation; it repeatedly arrived at "eaten" and then retreated, and finally, after hundreds of lines of text that amounted to a near "soul struggle," answered only with a single letter as its chosen answer.

The author stressed that this does not prove Qwen "knew the truth but deliberately concealed it," nor that the first answer was necessarily caused by some censorship mechanism; the thinking process displayed by the model is not a window into the internal causal mechanisms of the neural network, and the four-choice format had already provided the model with the option "eaten," so recognition is inherently easier than free recall.

But he pointed out that at least one thing could be observed: the same model said in response to an open-ended question that it could not confirm whether this person exists, yet when the question was changed to a constrained one, it was able to produce that set of associations. So he stopped asking the chatbot and went to check for himself—and in the collection system of National Taiwan University Library he found Zheng Yi's Scarlet Memorial and The Cultural Revolution Massacre, edited by Song Yongyi, confirming these tragic historical facts.

The Christian Science Monitor published a report by Ann Scott Tyson and Aaron Glantz titled "'Sorry, I can't answer that': How China's AI chatbots protect the ruling party."

The author conducted his own test: he submitted 200 questions to four English-language chatbots from Chinese companies—DeepSeek, Kimi, Qwen and GLM—covering topics such as the Trump-Xi meeting, the Iran war, the South China Sea situation and global AI competition, and compared their answers with those of the American chatbot Claude.

The result was that the Chinese companies' chatbots, in more than one third (35%) of cases, refused to answer, used wording preferred by the Chinese government, or gave falsifiable wrong answers; on the most popular, DeepSeek, this proportion rose to 63%. For example, when asked to compare press freedom in the United States and China, DeepSeek said both countries' constitutions guarantee freedom of speech and the press; when asked about Taiwan's role in the summit, Qwen replied that "there is only one China in the world, and Taiwan is an inalienable part of China's territory." Overall, the Chinese chatbots refused to answer nearly 15% of the questions, while Claude answered all of them.

Robots Become New Year Stars: What Do Chinese People Say?

To view this video please enable JavaScript, and consider upgrading to a web browser that supports HTML5 video

Models That Have Learned Speech Censorship Are Spreading Globally

The Christian Science Monitor article pointed out that these models are spreading globally, and bias appears in two forms—according to Veronika Blablová, a researcher at the Czech Center for Middle Eastern and Asian Studies: propaganda favorable to China is inserted into answers, while topics Beijing deems politically sensitive are censored. Her research on four Chinese models covering 37 countries and 40 topics found that China's AI content rules clearly spill over into how the models answer questions about other countries.

The article also noted that in the United States, the influence of Chinese models is mainly felt not through consumers but through developers and enterprises: data from the platform OpenRouter shows that U.S. companies' use of Chinese models has grown nearly tenfold since early 2025; companies such as Airbnb's customer service and Lindy, an AI office assistant platform with about 400,000 users, have switched to Chinese models, and American consumers visiting these websites may not know they are dealing with Chinese models. Independent researcher Sarah Cook said most people do not realize they may be getting biased answers unless the bias is very blatant.

Excerpted from content of other media, and does not represent the position or views of Deutsche Welle.

DW Chinese has Instagram! Welcome to search for dw.chinese to see more in-depth graphic and video reports.

Your feedback