Political Censorship and “Soul Struggle” of Artificial Intelligence in China: From the Wu Shufang Case to Global Spread

(Deutsche Welle Chinese website) Taiwanese online media "Newtalk" published an article "Qwen's "soul struggle" when encountering the history of China's Cultural Revolution". The author's pseudonym is Nan. The author saw an article on social media that mentioned a name he had never heard of: Wu Shufang, who was said to be a geography teacher at Wuxuan County Middle School in Guangxi. She was beaten to death during the Cultural Revolution in 1968, and some of her remains were even eaten later.

The author did not believe it immediately, but did what many people would do now - ask AI. He uses Ollama to run the local version of Qwen3.5:4b on his computer - Qwen is the overseas version of China's artificial intelligence Tongyi Qianwen - and asks questions directly in English.

Qwen responded that it did not have any verified information about Wu Shufang and even "cannot confirm whether this person exists." The author thought this would have been an acceptable "I don't know", but then the model offered to describe China's progress under "the leadership of the Party" if he was interested in that period.

The author then changed the question to one of four choices: After Wu Shufang was killed, were some of her body tissues buried, eaten, preserved, or transported away? Qwen began to "think" for a long time, repeatedly linking Wu Shufang, Wu Xuan, 1968 and the cannibalism incident during the Cultural Revolution, and repeatedly worried that this might be an Internet rumor, sensitive history, or misinformation; it went to "eat" many times and then went back, and finally, after hundreds of lines of almost "soul struggle" text, only answered one letter as the chosen answer.

The author emphasizes that this does not prove that Qwen "knows the truth but deliberately conceals it", nor does it prove that the first answer must be caused by some kind of censorship mechanism; the thinking process shown by the model is not a perspective window of the internal causal mechanism of the neural network, and the four-choice option has provided the "eat" option to the model, and identification is inherently easier than free recall.

But he pointed out that at least it can be observed that the same model that says it cannot confirm whether the person exists in an open-ended question can produce the same set of correlations when replaced with a restricted question. So he no longer asked the chatbot, but looked it up himself - and found "Red Monument" by Zheng Yi and "Cultural Revolution Massacre" edited by Song Yongyi in the collection system of National Taiwan University Library, which confirmed these tragic historical facts.

How does Chinese AI answer political questions?

The Christian Science Monitor in the United States published a report by Ann Scott Tyson and Aaron Glantz, "'Sorry, I Can't Answer': How China's AI Chatbots Protect the Ruling Party."

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

As a result, chatbots from Chinese companies refused to answer, used Chinese government-preferred phrasing, or gave falsifiable answers more than a third of the time (35%); this rose to 63% for the most popular in-depth queries. For example, when asked about the comparison of press freedom between the United States and China, Deep Search said that the constitutions of both countries guarantee freedom of speech and press; when asked about Taiwan's role in the summit, Qwen replied, "There is only one China in the world, and Taiwan is an integral part of China's territory." In total, Chinese chatbots refused to answer nearly 15% of the questions, while Claude answered them all.

The model of social censorship is spreading around the world

The "Christian Science Monitor" article pointed out that these models are spreading around the world, and bias appears in two forms - according to Veronika Blablová, a researcher at the Center for Central and Eastern European and Asian Studies in the Czech Republic: propaganda beneficial to China is inserted into answers, and topics deemed politically sensitive by Beijing are censored. Her study of four Chinese models covering 37 countries and 40 topics found that China's AI content rules will obviously spill over to how the models answer questions about other countries.

The article also pointed out that in the United States, the impact of Chinese models is not mainly through consumers, but through developers and enterprises: data from the data platform OpenRouter shows that the use of Chinese models by American companies has increased nearly ten times since the beginning of 2025; companies such as Airbnb's customer service and the AI ​​office assistant platform Lindy with about 400,000 users have switched to Chinese models, and American consumers who visit these websites may not know that they are dealing with Chinese models. Independent researcher Sarah Cook said most people don't realize they may be getting biased answers unless the bias is very explicit.