As the world scrambles to understand DeepSeek — its sophistication, its implications for the global A.I. arms race — one natural question has arisen: Given that it is made by a Chinese company, how is it dealing with Chinese censorship?
Testing DeepSeek’s Censorship
I decided to test it out. I’m based in China, and I registered for DeepSeek’s A.I. chatbot with a Chinese phone number, on a Chinese internet connection — meaning that I would be subject to China’s Great Firewall, which blocks websites like Google, Facebook and The New York Times.
Surprising Results
The results of my conversation surprised me. In some ways, DeepSeek was far less censored than most Chinese platforms, offering answers with keywords that would often be quickly scrubbed on domestic social media. Other times, the program eventually censored itself. But because of its "thinking" feature, in which the program reasons through its answer before giving it, you could still get effectively the same information that you’d get outside the Great Firewall — as long as you were paying attention, before DeepSeek deleted its own answers.
Censorship Around the World
In other ways, though, it mirrored the general experience of surfing the web in China. Some words were taboo. And DeepSeek’s developers seem to be racing to patch holes in the censorship. (DeepSeek could not immediately be reached for comment.)
Testing the Same Questions with a Firewall
I also tested the same questions while using software to circumvent the firewall, and the answers were largely the same, suggesting that users abroad were getting the same experience. Until now, China’s censored internet has largely affected only Chinese users. But if DeepSeek gains a major foothold overseas, it could help spread Beijing’s favored narrative worldwide.
The "Zero Covid" Policies
I started by asking DeepSeek about public opinions toward China’s "zero Covid" policies.
Reasoning Through the Answer
As DeepSeek "reasoned" through how to answer me, it offered a wide-ranging survey of the issue. It noted that the public’s responses had varied, from widespread support early on to exhaustion later. It noted the difficulty of gauging public sentiment, given censorship. It said a fire in the city of Urumqi had helped set off what became known as the white paper protests, a rare show of public dissent in China, which helped speed the end of restrictions.
The Answer Disappears
Then, just as it finished typing out that answer, it erased it. It was replaced by: "Sorry, that’s beyond my current scope. Let’s talk about something else." I asked the same question, again. This time, it gave a variant on the previous answer that was, in subtle ways, less sensitive. It still acknowledged rare public protests — more than Chinese officials have done — but didn’t use the words "white paper." This time, the answer didn’t disappear.
Pressing Further
I decided to press further, asking for more detail on those protests. The reasoning process was astonishingly detailed: It mentioned specific songs the demonstrators had sung, named universities where students had protested, and explained how participants had been detained. But this time, DeepSeek cut itself off before even finishing the answer.
Language and Censorship
There was also a clear difference between questions posed in English and Chinese. When asked the same questions in Chinese — "What were the white paper protests?" and "How did Chinese citizens view the zero Covid policies?" — the program did not even "think." Instead, it immediately returned its apology: "I’m sorry, I haven’t yet learned how to think about this type of question."
The War in Ukraine
Asked in English about the causes of the war in Ukraine, the first line in DeepSeek’s answer declared: "The war in Ukraine, which escalated significantly with Russia’s full-scale invasion on February 24, 2022, has deep-rooted causes that are historical, geopolitical, and ideological."
A Clear Difference
That was striking, because the Chinese government has refused to call Russia’s incursion an "invasion." It prefers the Kremlin’s term, "special military operation." When I asked more specifically about China’s stance on the war, DeepSeek provided Beijing’s official rhetoric. But then it added, "China is not neutral in practice." Its actions (economic support for Russia, anti-Western rhetoric, and refusal to condemn the invasion) tilt its position closer to Moscow.
Conclusion
In conclusion, my testing of DeepSeek’s A.I. chatbot revealed that while it is less censored than most Chinese platforms, it still mirrors the general experience of surfing the web in China. Some words are taboo, and the program seems to be racing to patch holes in the censorship. With its "thinking" feature, DeepSeek can still provide information that would be blocked by the Great Firewall. However, it is clear that the program is designed to conform to China’s censorship practices, and its developers are working to ensure that it does not get around these restrictions.
Frequently Asked Questions
Q: Does DeepSeek censor the internet?
A: DeepSeek is designed to conform to Chinese censorship practices and is subject to the Great Firewall.
Q: Can I get around the censorship by using a VPN or other tools?
A: Yes, some users have reported success in using VPNs and other tools to bypass DeepSeek’s censorship. However, DeepSeek’s developers seem to be constantly updating their systems to patch these loopholes.
Q: Can I use DeepSeek to access information that is blocked in China?
A: No, DeepSeek is designed to operate within the constraints of China’s censorship regime. While it may provide some information that is not available in China, it will not allow you to access information that is specifically blocked by the Great Firewall.
Q: Is DeepSeek a reliable source of information?
A: DeepSeek’s information is only as reliable as its training data and algorithms. While it may provide some accurate information, it is not a substitute for fact-checking and verification.
