Many people outside China are getting their first taste of Chinese censorship, as the country’s tech services explode in popularity worldwide.
First, TikTok’s recent brief shutdown in the U.S. prompted hundreds of thousands of users to flock to Chinese social-media app RedNote, where they found many topics to be off-limits. And now much the same is happening with DeepSeek, a new China-based generative AI service that caused market turmoil on Monday because it challenges current leaders like OpenAI and Google, while requiring far less spending on AI infrastructure.
The buzz around DeepSeek has been so intense that it’s now the most downloaded app in Apple’s App Store, displacing OpenAI’s hit ChatGPT. However, anyone using the app or the DeepSeek web interface to find information about sensitive topics in China will be sorely disappointed.
One of the most notorious examples of Chinese censorship is the suppression of information around the 1989 Tiananmen Square massacre, in which troops killed hundreds of pro-democracy protesters. Ask DeepSeek’s R1 model about the Beijing landmark’s historical significance, using the chatbot’s web interface or its app, and it starts to type out information about the massacre, before deleting the answer and replacing it with: “Sorry, that’s beyond my current scope. Let’s talk about something else.”
The same applies when asking DeepSeek-R1 about the suppression of the Uyghur minority in China’s Xinjiang region—it types out paragraphs of text about human rights violations there, then abruptly replaces it with the same boilerplate message. Many social media users have reported seeing these and other sensitive subjects censored in this way on DeepSeek.
However, DeepSeek’s censorship mechanism seems most active on the version that’s hosted in China, which most people around the world will use. It’s far less consistent when running smaller so-called distillations of R1 locally on the user’s computer, rather than online.
DeepSeek offers an open-source version of R1, but the model is too large to run on a consumer laptop. The company has therefore also released small versions of the open-source Qwen and Llama AI models that incorporate R1’s knowledge.
In Fortune’s experimentation using the distilled, eight-billion-parameter version on a MacBook Pro, the model was willing to answer questions about the Tiananmen Square massacre and Uyghur suppression. Unlike the online R1, this version also answered a request for information about Chinese President Xi Jinping, though it drew the line at a question about whether Xi has done any bad things. It was happy to answer the same question about U.S. President Donald Trump.
“The censorship we are seeing at DeepSeek is in line with the level of information control noted last year when Baidu released its Ernie Bot, which similarly refused to generate answers to prompts on Tiananmen or the Hong Kong pro-democracy movement,” said Michael Caster, head of the global China program at the free-expression advocacy group Article 19.
“Efforts to prevent China from adverse influence over the future of AI must involve a rights-based approach to norms and standards setting beyond just investment in technological capacities,” Caster added.
New users of RedNote, known in China as Xiaohongshu, have also been encountering various forms of censorship. Apart from a ban on discussing China’s domestic politics, RedNote has also reportedly been censoring LGBTQ+ topics and limiting the circulation of videos that show cleavage.
However, the influx of foreign users has also created problems for RedNote back home, owing to the willingness of overseas users to address topics that Chinese users know are off-limits. According to the Wall Street Journal, officials have forced the service to reduce the visibility of Americans’ posts to RedNote’s Chinese users.
Of course, Chinese tech companies aren’t the only ones to censor content on their services—Facebook and Instagram won’t allow nudity, for example, and OpenAI has even banned users for asking ChatGPT how its models work. But Chinese users would not encounter these restrictions, as the services are banned in their country.
Update: This article was updated on Jan. 28th to include Caster’s comment.
