Leaked AI Dataset Reveals China’s Censorship Ambitions

A leaked dataset has revealed how Chinese entities are training large language models (LLMs) to automate political censorship on a massive scale. The dataset, containing over 133,000 real-world content examples, includes posts about government corruption, rural poverty, Taiwan, and military affairs, topics the Chinese state typically considers sensitive.

According to TechCrunch, the system is designed to flag this content automatically, providing a glimpse into how artificial intelligence is being deployed to refine and scale digital repression. UC Berkeley researcher Xiao Qiang, who examined the dataset, argued that this is clear evidence that the Chinese government or its affiliates want to use LLMs to improve repression.

Unlike older systems that relied on keyword filters and human moderation, this LLM-based approach enables more efficient control over online discourse.

Security researcher NetAskari discovered the unsecured database on a Baidu server, and found that it contained entries as recent as December 2024. Though the creators are unidentified, the dataset is marked for “public opinion work”, a term widely associated with censorship operations led by the Cyberspace Administration of China.

While the TechCrunch report does not name a specific model, separate investigations suggest that DeepSeek AI, one of China’s most prominent open-source LLMs, is already exhibiting censorship behaviors consistent with the leaked system’s goals.

WIRED tested DeepSeek-R1 across platforms and found that the model censors topics like Taiwan and Tiananmen through both app-level filtering and pre-programmed bias. In one case, the model’s internal reasoning noted the need to “avoid mentioning events that could be sensitive,” while emphasizing China’s achievements under the Communist Party.

Adding to the long list of concerns, Feroot Security recently found that DeepSeek’s platform contains hidden code transmitting user data to servers controlled by China Mobile, a state-owned telecom company under US sanctions.

As AI tools become more embedded in everyday platforms, experts warn that state-aligned models could shape global information flows. Several countries, including the US, Italy, and Australia, are now evaluating bans or restrictions on Chinese AI systems. The rise of censorship-enabled LLMs, researchers say, marks a turning point in how digital authoritarianism is executed.