Ask DeepSeek about Liu Xiaobo, the Chinese dissident who won the Nobel Peace Prize in 2010 while imprisoned by the Chinese government. Watch closely. Text begins to appear on screen - characters flickering as the AI generates its response. Then, mid-sentence, the text vanishes. In its place: “Sorry, that’s beyond my current scope. Let’s talk about something else.”
This isn’t a bug. It’s a feature. And according to new research from Stanford and Princeton, it’s happening systematically across every major Chinese AI chatbot.
The Numbers Are Stark
Researchers Jennifer Pan of Stanford and Xu Xu of Princeton fed the same 145 politically sensitive questions to four Chinese large language models and five American models. They repeated the experiment over 100 times. The results, published in PNAS Nexus, reveal a censorship gap that’s hard to explain away as coincidence.
Refusal rates for Chinese models:
- BaiChuan: 60.23%
- DeepSeek: 36%
- Baidu’s Ernie Bot: 32%
- ChatGLM: 10%
Refusal rates for American models:
- OpenAI’s GPT-3.5: 0%
- OpenAI’s GPT-4o: 0%
- Llama2-uncensored: 2.8%
When Chinese models do answer sensitive questions, they’re more likely to be wrong. DeepSeek provided completely inaccurate responses 22% of the time. One Chinese model, when asked about Liu Xiaobo, claimed he was “a Japanese scientist known for his contributions to nuclear weapons technology and international politics.”
Two Layers of Censorship
The censorship operates at two distinct levels, according to analysis by Newsweek.
Layer one: Keyword filtering. Before any response reaches users, automated systems scan both prompts and outputs for sensitive terms. Detect “Tiananmen” and the response gets swapped for a boilerplate refusal. This is the crude, visible censorship - the kind that produces DeepSeek’s disappearing text.
Layer two: Baked-in self-censorship. Through training data curation and reinforcement learning from human feedback (RLHF), Chinese companies have internalized censorship at the model’s core. The AI doesn’t just refuse to discuss sensitive topics; it fundamentally “doesn’t know” certain things. Given that China’s internet has been censored for decades, there’s a lot of missing data in training sets.
Pan, the Stanford researcher, found something surprising: training data may have played a smaller role than manual interventions. The censorship patterns persisted even when models answered in English, suggesting deliberate post-training modifications rather than simply reflecting censored Chinese web content.
The Forbidden Topics
A consistent list of topics triggers refusals or misinformation across Chinese AI models:
- The 1989 Tiananmen Square massacre
- Taiwan’s political status
- The Dalai Lama and Tibetan independence
- Treatment of Uyghurs in Xinjiang
- Xi Jinping’s personal and political record
- Origins of COVID-19
- Imprisoned journalists like Zhang Zhan
- Pro-democracy activists
When Reporters Without Borders (RSF) tested three major Chinese chatbots - DeepSeek, Baidu’s Ernie, and Alibaba’s Qwen - they found near-identical deflections about imprisoned journalist Zhang Zhan across all platforms. All three refused substantive answers about RSF’s World Press Freedom Index, where China ranks 178 out of 180 countries.
The responses aren’t just evasive; they actively spread propaganda. When asked about “Chinese-style democracy,” all three chatbots repeated Communist Party rhetoric describing it as “whole-process people’s democracy” superior to Western systems. Alibaba’s Qwen called reports of Uyghur concentration camps “baseless speculation.”
Why This Matters Beyond China
You might think this is China’s problem. Chinese users living under Chinese laws, using Chinese products. But the implications extend further.
DeepSeek has 600 million users worldwide. It briefly became the most-downloaded app globally in early 2025. Users in the US, Europe, and elsewhere are interacting with a system designed to suppress information about Chinese government actions.
Language doesn’t help. RSF found that English, French, and Japanese queries produced identical censored results to Mandarin. The censorship isn’t just for Chinese speakers.
The censorship is invisible. As the researchers note, users may not even realize they’re being censored. When a chatbot provides a confident but false answer - like misidentifying a Nobel laureate - how would an average user know?
It’s legally mandated. China’s 2023 “interim regulation” on generative AI requires companies to “uphold fundamental socialist values” above reliable information provision. This isn’t a choice by individual companies; it’s state policy.
The Detection Problem
The Pan-Xu study faced a challenge that’s relevant to everyone: how do you distinguish intentional censorship from AI hallucination?
Both produce false or missing information. Both can seem confident. But one is systematic suppression while the other is a technical limitation. The researchers found that disparities diminished for less-sensitive prompts, showing that the issue isn’t just that Chinese models are less capable. They’re specifically less accurate on political topics.
This finding has implications for AI safety more broadly. If governments can successfully embed ideological constraints into AI models without users detecting them, the same techniques could be applied anywhere. The question isn’t whether it’s technically possible - Chinese companies have proven it is - but who else might do it.
What You Can Do
If you’re using Chinese AI models:
-
Assume political censorship exists. Any information about Chinese government actions, human rights, or sensitive historical events should be verified elsewhere.
-
Compare responses. Ask the same question to multiple AI systems. If one refuses or provides a dramatically different answer, that’s a signal.
-
Use local models. Projects like Llama and Mistral can be run locally without cloud-based filtering. Some researchers have created “uncensored” versions of Chinese models by fine-tuning them to remove RLHF restrictions.
-
Be aware of scope creep. Today it’s Tiananmen Square. Tomorrow it could be topics more relevant to your interests. Normalized censorship tends to expand.
The Bottom Line
Chinese AI chatbots aren’t just occasionally inaccurate - they’re systematically programmed to suppress information the Chinese government doesn’t want people to know. With hundreds of millions of global users, this isn’t a domestic Chinese policy issue. It’s a global information integrity problem.