Chinese AI chatbots censor details of Mbappé's Mao Zedong remarks

French football star Kylian Mbappé sparked significant diplomatic friction with China after a mid-October podcast appearance. During an episode of The Bridge, the Real Madrid striker referenced a popular meme featuring his head superimposed on the body of a figure he identified as a Chinese dictator. This reference was widely understood to be a joke about Mao Zedong, the founder of the People's Republic of China. The comment quickly led to accusations that Mbappé had insulted the nation's national sentiments.

The controversy prompted immediate commercial and diplomatic reactions. Major Chinese sponsors, including beverage giant Dongpeng Group, dairy company Mengniu, and cheesemaker Milkground, severed their partnerships with the player. A spokesperson for the Chinese Ministry of Foreign Affairs also criticized Mbappé during a press conference, although the footballer has since issued an apology that officials say is being taken seriously.

Beyond the human reaction, a digital layer of censorship has emerged. A Chinese dissident account on X first noted that domestic AI chatbots were refusing to provide information about the incident. The France 24 Observers team subsequently tested several AI models to understand how these systems handled queries regarding the dispute. The tests focused on DeepSeek, a high-performing Chinese model, and Doubao, an AI developed by ByteDance for users within China. They also tested Dola, the version of ByteDance's AI provided to international users.

To simulate conditions for users inside China, the team used a virtual private network to access the chatbots with a Chinese IP address. When asked in Chinese if Mbappé had called Mao Zedong a dictator, Doubao responded that the content was suspected of violating its usage rules. DeepSeek suggested the story was likely fake news, stating there was no trustworthy public information to confirm the statement and that the footballer rarely discusses political topics.

The behavior of these models changed when accessed from outside China. Using a French IP address, DeepSeek still refused to answer the specific question, claiming the response went beyond its remit. However, Dola, the international version of ByteDance's tool, acknowledged that Mbappé alluded to Mao Zedong being a dictator. When asked the same question in Chinese via Dola, the response shifted to state that public opinion believed Mbappé was referring to the former leader.

When asked more indirectly what Mbappé exactly said, Chinese IP-based models remained vague. They condemned the footballer for making inappropriate statements that wounded the Chinese people but avoided repeating the specific words or naming Mao Zedong. Doubao cited a zero-tolerance policy for acts harming national sentiment, while DeepSeek noted that media outlets had deliberately chosen not to translate or disseminate the exact wording of the controversial remarks.

Expert analysis suggests this censorship is tied to the political legacy of Mao Zedong. Marie Holzman, an expert on China, explained that Mao's image is crucial for President Xi Jinping's legitimacy. Criticizing Mao can cast doubt on the Chinese Communist Party's historical narrative. Jennifer Pan, a professor of political communication at Stanford University, added that the reaction fits a broader narrative of the West disrespecting China's rise. She noted that while Mao is not a taboo, versions of his history that differ from the official party line are often suppressed.

Current regulations in China require AI to promote fundamental societal values and ban content that incites subversion or tarnishes the nation's image. Research indicates that Chinese AI models are more likely to give short, vague, or refusing responses to political questions compared to models developed elsewhere. This suggests that censorship is integrated into the training and operational protocols of these domestic systems.