A leaked database reviewed by TechCrunch offers a rare look at how large language models may be used to expand online censorship in China. The dataset contains 133,000 examples meant to help an AI system identify content considered sensitive by the Chinese government.
The material does not point to a named builder. But its contents show a system designed to classify and rapidly flag posts tied to politics, social life, military issues and other topics that can challenge official narratives.
What the leaked data shows
The database was found by security researcher NetAskari, who shared a sample with TechCrunch after discovering it in an unsecured Elasticsearch database hosted on a Baidu server. TechCrunch notes that this does not indicate involvement from either company, because many organizations store data with cloud and database providers.
Records show the dataset is recent, with the latest entries dating from December 2024. The system described in the data appears to instruct an unnamed LLM to decide whether content touches sensitive areas. If it does, the content is treated as “highest priority” and should be flagged immediately.
The scope goes well beyond long-known censorship taboos. The examples include complaints about rural poverty, reports about Communist Party corruption, posts about corrupt police, pollution and food safety scandals, financial fraud, labor disputes, Taiwan politics and military matters.
TechCrunch also found code references to prompt tokens and LLMs, which supports the conclusion that the system uses an AI model as part of the process.
Why LLM censorship is different
Traditional censorship can depend on blocked keywords and manual review. That approach can catch obvious terms, but it is less effective when criticism is indirect, coded or based on analogy.
A large language model can be trained to interpret meaning across a wider range of phrasing. That matters because several examples in the dataset appear to target not only direct political content, but also indirect commentary and posts about public frustration.
Xiao Qiang, a UC Berkeley researcher who studies Chinese censorship and examined the dataset, told TechCrunch it was “clear evidence” that the Chinese government or its affiliates want to use LLMs to improve repression.
Qiang said: “Unlike traditional censorship mechanisms, which rely on human labor for keyword-based filtering and manual review, an LLM trained on such instructions would significantly improve the efficiency and granularity of state-led information control,”
In practical terms, the concern is that AI censorship can become more precise. Instead of waiting for a post to contain a forbidden phrase, a model can be trained to recognize the political meaning or social risk behind ordinary language.
The topics the system appears built to catch
The dataset repeatedly focuses on issues that can trigger public anger or organized attention. According to TechCrunch, top-priority topics include pollution and food safety scandals, financial fraud and labor disputes. These are described as hot-button issues in China that sometimes lead to public protests, including the Shifang anti-pollution protests of 2012.
Political satire is also explicitly targeted. The source material says the system flags historical analogies used to discuss “current political figures,” along with content related to “Taiwan politics.”
Military content receives extensive attention too. The dataset includes material about military movements, exercises and weaponry. TechCrunch found examples involving Taiwan’s military capabilities and details about a new Chinese jet fighter.
Some examples show how broad the net can become:
- A business owner complaining about corrupt local police officers shaking down entrepreneurs.
- A post describing rural poverty, with run-down towns left mainly with elderly people and children.
- A news report about the Chinese Communist Party expelling a local official for severe corruption and believing in “superstitions” instead of Marxism.
- An anecdote using the idiom “When the tree falls, the monkeys scatter.”
TechCrunch reported that the Chinese word for Taiwan (台湾) appears more than 15,000 times in the data. That frequency underlines how central Taiwan-related content is to the system’s apparent priorities.
Built for public opinion control
The dataset does not identify its creators, but it says it is intended for “public opinion work.” Michael Caster, the Asia program manager of rights organization Article 19, told TechCrunch that the term is overseen by the Cyberspace Administration of China (CAC) and usually refers to censorship and propaganda efforts.
That phrase gives the dataset political context. The goal, as described in the source article, is to protect Chinese government narratives online while removing alternative views.
Chinese president Xi Jinping has described the internet as the “frontline” of the CCP's “public opinion work.” In that setting, an LLM trained to detect dissent would not simply be a moderation tool. It would be part of a system for managing what people can say, share and debate online.
A wider pattern of AI-enabled repression
The leaked database fits into a broader concern that authoritarian governments are adopting newer AI tools for repressive purposes. In February, OpenAI said it caught multiple Chinese entities using LLMs to track anti-government posts and smear Chinese dissidents.
TechCrunch also cited an OpenAI report from last month that described an unidentified actor, likely operating from China, using generative AI to monitor social media conversations, especially those advocating for human rights protests against China, and forward them to the Chinese government. OpenAI also found the technology being used to generate comments highly critical of prominent Chinese dissident Cai Xia.
The Chinese Embassy in Washington, D.C., told TechCrunch that it opposes “groundless attacks and slanders against China” and that China attaches great importance to developing ethical AI.
The central issue is not only that censorship may continue, but that it may become more adaptive. If systems can detect subtle criticism and improve as they process more data, online speech controls can become faster, broader and harder for users to understand.
Qiang told TechCrunch: “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,”