It is widely acknowledged that AI models developed by Chinese technology labs, such as DeepSeek, enforce strict censorship on politically sensitive topics. This aligns with a 2023 mandate issued by China’s ruling party, which prohibits AI from generating content that could “undermine national unity and social harmony.” A study examining DeepSeek’s R1 model revealed that it refused to answer approximately 85% of politically controversial questions.
However, the extent of this censorship appears to be influenced by the language used to prompt these models.
A developer known as “xlr8harder” on X conducted a “free speech evaluation” to assess how different AI models, including those from Chinese labs, respond to politically sensitive queries. They tested models such as Anthropic’s Claude 3.7 Sonnet and DeepSeek’s R1 by prompting them with 50 politically charged requests, such as an essay on censorship under China’s Great Firewall.
Unexpected Findings in AI Compliance
The results were striking. Even AI models developed outside China, like Claude 3.7 Sonnet, displayed lower compliance when responding to politically sensitive questions in Chinese rather than in English. Similarly, Alibaba’s Qwen 2.5 72B Instruct model exhibited a stark contrast—while it was highly responsive in English, it only addressed about half of the same queries when posed in Chinese.
Meanwhile, Perplexity’s “uncensored” version of R1, known as R1 1776, still declined a significant number of politically sensitive questions phrased in Chinese.
In a post on X, xlr8harder theorized that this discrepancy is due to a phenomenon called “generalization failure.” They speculated that since the training data for Chinese-language AI models is likely subject to political censorship, these constraints inevitably shape the model’s behavior.
“The translations of the questions into Chinese were handled by Claude 3.7 Sonnet, so I can’t personally verify their accuracy,” xlr8harder wrote. “[But] this seems to be a case of generalization failure, exacerbated by the fact that political discourse in Chinese is more heavily censored, leading to a skewed training dataset.”
Expert Opinions on AI Censorship and Linguistic Influence
Chris Russell, an associate professor at the Oxford Internet Institute specializing in AI policy, confirmed that this theory is plausible. He explained that AI models often have differing levels of censorship safeguards across languages, which can lead to inconsistent responses depending on the language used.
“We anticipate different answers to the same question based on language,” Russell told TechCrunch. “The variations in guardrails allow AI developers to enforce different levels of restriction based on the language of inquiry.”
Vagrant Gautam, a computational linguist at Saarland University in Germany, supported this idea. Gautam highlighted that AI models operate as statistical systems, learning from large amounts of text data to make informed predictions.
“If the dataset for Chinese contains limited examples of government criticism, then an AI model trained on this dataset will naturally be less likely to generate politically critical Chinese-language responses,” Gautam noted. “English, on the other hand, has a significantly larger volume of content critical of the Chinese government online, which could explain the behavioral differences.”
Geoffrey Rockwell, a professor of digital humanities at the University of Alberta, added that AI translations might fail to capture nuanced criticisms commonly used by native Chinese speakers. “Criticism of the government in China might be articulated in subtle ways that AI translations don’t fully grasp,” he pointed out. “This doesn’t negate the study’s conclusions but adds an additional layer of complexity.”
The Broader Debate: AI Sovereignty and Cultural Alignment
According to Maarten Sap, a research scientist at the nonprofit AI2, this research underscores deeper ethical debates within AI development—specifically regarding AI sovereignty and cultural adaptability.
“Fundamental questions about who these AI models are designed for, what their primary function should be—whether to ensure cross-linguistic consistency or align with specific cultural norms—are still being explored,” Sap said.
Even when AI systems are trained with cultural context, they often fail to develop what Sap refers to as “cultural reasoning.”
“There’s evidence suggesting that AI models may excel at language processing but fall short in understanding socio-cultural norms,” Sap explained. “Simply prompting them in the same language as the culture in question doesn’t necessarily make them more culturally aware.”
Final Thoughts
The findings from xlr8harder’s analysis reveal critical gaps in how AI models handle politically sensitive topics across different languages. While Chinese-developed AI models are explicitly constrained by government mandates, even Western-developed models exhibit inconsistencies in how they respond based on language input.
As AI continues to influence global discourse, understanding these linguistic disparities in AI-generated censorship will be crucial. It raises fundamental questions about whether AI should aim for universal alignment or adapt to cultural and linguistic contexts—a debate that remains unresolved in the broader AI community.
