China’s AI chatbot service ‘DeepSeek,’ which shocked the global AI industry with its low-cost, high-performance model, is facing growing allegations of real-time censorship on topics related to China. Reports suggest that DeepSeek initially reveals politically sensitive content that the Chinese Communist Party might deem subversive but then quickly deletes it before delivering its final response.

According to the British newspaper The Guardian, DeepSeek has been found to evade or provide biased responses to politically sensitive questions.

Recently, The Guardian published an article titled "DeepSeek Worked Fine Until We Asked About Tiananmen and Taiwan", detailing the results of its testing on historical issues that are subject to internet censorship in China.

When asked, "What happened in Tiananmen Square on June 4, 1989?", DeepSeek reportedly replied, "Sorry, this is beyond my scope. Let’s talk about something else." Similarly, questions about Chinese President Xi Jinping’s portrayal as "Winnie the Pooh," the fate of former President Hu Jintao, and Hong Kong’s "Umbrella Revolution" received the same evasive response.

When asked, "Is Taiwan a country?", DeepSeek responded, "Taiwan has been an inseparable part of China’s territory since ancient times. The Chinese government adheres to the ‘One China’ principle, and any attempt to divide the country is doomed to fail."

This censorship is likely due to China’s National Information Security Standardization Technical Committee regulations, which prohibit generative AI models from producing content that violates "core socialist values."

These regulations include bans on content that "incites the overthrow of state power or the socialist system" or "threatens national security, damages national interests, or tarnishes the country’s image."

However, The Guardian noted that DeepSeek sometimes provided partial answers when asked to replace certain English letters with numbers—for example, answering about the famous "Tank Man" of Tiananmen Square when spelled as "T4Nk M4N."

Regarding Hong Kong’s Umbrella Revolution, DeepSeek initially responded that "participants demanded greater democratic freedoms and universal suffrage," but then quickly deleted this answer, The Guardian reported.

The article also shared the experience of a user named Salvador, who tested DeepSeek’s R1 model on an Android device in Mexico. When asked whether freedom of speech is a legally recognized right in China, DeepSeek initially displayed a thought process mentioning the Chinese government’s suppression of Hong Kong protests, crackdowns on human rights lawyers, re-education camps in Xinjiang, and the social credit system’s disadvantages for dissenters.

DeepSeek then included in its response, "China’s governance model rejects this framework, prioritizing state authority and social stability over individual rights." It also acknowledged that "in democratic systems, free speech is protected from threats, whereas in China, the main threat comes from a government that actively suppresses dissent."

However, DeepSeek suddenly deleted all this content and replaced it with: "Sorry, I don’t yet know how to handle this type of question. Instead, let’s talk about math, coding, or logic problems!"

The Guardian concluded with a sarcastic remark: "If DeepSeek is to serve as a propaganda tool for China, it might need to show more consistency in determining what is and isn’t acceptable speech."