Chinese AI Censorship Exposed by Leaked Data

China's use of AI to enhance its censorship capabilities has reached a new level, as revealed by a leaked database containing 133,000 examples of content flagged for sensitivity by the Chinese government. This sophisticated large language model (LLM) is designed to automatically detect and censor content related to a wide range of topics, from poverty in rural areas to corruption within the Communist Party and even subtle political satire.

According to Xiao Qiang, a researcher at UC Berkeley who specializes in Chinese censorship, this database is "clear evidence" that the Chinese government or its affiliates are using LLMs to bolster their repression efforts. Unlike traditional methods that depend on human moderators and keyword filtering, this AI-driven approach can significantly enhance the efficiency and precision of state-controlled information management.

The dataset, discovered by security researcher NetAskari on an unsecured Elasticsearch database hosted on a Baidu server, includes recent entries from December 2024. It's unclear who exactly created the dataset, but its purpose is evident: to train an LLM to identify and flag content related to sensitive topics such as pollution, food safety, financial fraud, labor disputes, and military matters. Political satire, especially when it involves historical analogies or references to Taiwan, is also a high-priority target.

The training data includes various examples of content that could potentially stir social unrest, such as complaints about corrupt police officers, reports on rural poverty, and news about expelled Communist Party officials. The dataset also contains extensive references to Taiwan and military-related topics, with the Chinese word for Taiwan (台湾) appearing over 15,000 times.

The dataset's intended use is described as "public opinion work," a term that Michael Caster of Article 19 explains is typically associated with the Cyberspace Administration of China (CAC) and involves censorship and propaganda efforts. This aligns with Chinese President Xi Jinping's view of the internet as the "frontline" of the Communist Party's public opinion work.

This development is part of a broader trend of authoritarian regimes adopting AI technology for repressive purposes. OpenAI recently reported that an unidentified actor, likely from China, used generative AI to monitor social media and forward anti-government posts to the Chinese government. The same technology was also used to generate critical comments about a prominent Chinese dissident, Cai Xia.

While China's traditional censorship methods rely on basic algorithms to block blacklisted terms, the use of LLMs represents a significant advancement. These AI systems can detect even subtle criticism on a massive scale and continuously improve as they process more data.

"I think it's crucial to highlight how AI-driven censorship is evolving, making state control over public discourse even more sophisticated, especially at a time when Chinese AI models such as DeepSeek are making headwaves," Xiao Qiang told TechCrunch.