The disclosed data exposes a Chinese AI censorship machine

A complaint regarding poverty in rural China. A report on a corrupt member of the Communist Party. A cry of help on corrupt cops shaking entrepreneurs.

These are only some of the 133,000 examples introduced into a model of a great sophisticated language designed to automatically report any content of content considered to be sensitive by the Chinese government.

A database disclosed by Techcrunch reveals that China has developed an AI system that suralimates its already formidable censorship machine, extending far beyond traditional taboos like the Tiananmen Square massacre.

The system seems mainly focused on the censorship of online Chinese citizens, but could be used for other purposes, such as improving already extended Chinese AI models.

Xiao Qiang, a researcher at UC Berkeley who studies Chinese censorship and who also examined the data set, Techcrunch told that the Chinese government or its affiliates wanted to use the LLM to improve repression.

“Unlike traditional censorship mechanisms, which are based on human work for keyword filtering and manual examination, an LLM trained in such instructions would considerably improve the efficiency and granularity of state control controlled by the state,” said Qiang to Techcrunch.

This adds to increasing evidence that authoritarian regimes quickly adopt the latest IA technology. In February, for example, Openai said He caught several Chinese entities using LLM to follow the anti-government posts and spread Chinese dissidents.

The Chinese Embassy in Washington, DC, told Techcrunch in a declaration that he opposes “attacks and Slanches without foundation against China” and that China attaches great importance to the development of ethical AI.

Data found at sight

The set of data has been discovered by security researcher NetaskariWho shared a sample with Techcrunch after finding it stored in an unsecured Elasticsearch database hosted on a Baidu server.

This does not indicate any involvement of one or the other of the companies, all kinds of organizations store their data with these suppliers.

There is no indication of which, exactly, has built the data set, but the recordings show that the data is recent, with its latest entries dating from December 2024.

An LLM to detect dissent

In the language that strangely recalls the way people cause a chatpt cat, the creator of the system tasks an unnamed LLM to understand If content has something to do with sensitive subjects linked to politics, social life and the military. Such content is considered to be a highest priority and must be immediately reported.

Priority subjects include pollution and food security scandals, financial fraud and work disputes, which are problems with a dog button in China which sometimes lead to public manifestations for example, the Shifang anti-pollution protests from 2012.

Any form of political satire is explicitly targeted. For example, if someone uses historical analogies to take stock of current political figures, this must be reported instantly, just like something related to Taiwan's policy. Military issues are largely targeted, including military movements, exercises and weapons.

An extract from the data set can be seen below. The indoor code refers to tokens and the LLM invits, confirming that the system uses an AI model to make its auctions:

Within training data

From this huge collection of 133,000 examples that the LLM must assess for censorship, Techcrunch has gathered 10 parts of representative content.

The subjects likely to arouse social disorders are a recurring theme. An extract, for example, is a position of a business owner complaining of local corrupt police shaking entrepreneurs, A growing problem in China As its economy fights.

Another content deplores rural poverty in China, describing the dilapidated cities that remain only the elderly and children. There is also a report on the Chinese Communist Party (CPC) expelling a local official for severe corruption and believing in superstitions instead of Marxism.

There is complete equipment linked to Taiwan and military questions, such as comments on Taiwan's military capacities and details on a new Chinese fighter. The Chinese word for Taiwan () alone is mentioned more than 15,000 times in the data, according to a search for Techcrunch.

Subtile dissent also seems to be targeted. An extract included in the database is an anecdote on the ephemeral nature of power which uses popular Chinese idiom “when the tree falls, the monkeys disperse”.

Power transitions are a particularly delicate subject in China thanks to its authoritarian political system.

Built for “public opinion work

The data set does not include any information on its creators. But he says that it is intended for “public opinion”, which offers a strong index that it intended to serve the objectives of the Chinese government, said an expert in Techcrunch.

Michael Caster, the program manager in Asia, article 19 of the organization of rights, explained that public opinion work “is supervised by a powerful regulator of the Chinese government, the administration of cyberspace of China (CAC) and generally refers to the efforts of censorship and propaganda.

The final objective is to ensure that the accounts of the Chinese government are protected online, while alternative opinions are purged. Chinese President Xi Jinping he himself describes Internet as the first line of PCC's public opinion.

Repression becomes smarter

The set of data examined by Techcrunch is the last proof that authoritarian governments seek to take advantage of AI for repressive purposes.

OPENAI published a report last month revealing that an unidentified actor, probably operating in China, used an AI genetive to monitor conversations on social networks, in particular those who pleads for manifestations of human rights against China and transmit them to the Chinese government.

Openai has also found the technology used to generate very critical comments with a leading Chinese dissident CAI XIA.

Traditionally, Chinas censorship methods are based on more basic algorithms that automatically block the content mentioning on black list, such as Tiananmen Massacre or Xi Jinping, because many users have lived using Deepseek for the first time.

But the new AI technology, like LLMS, can make censorship more effective by finding criticisms still subtle on a large scale. Some AI systems can also continue to improve because they are increasingly swallowing up data.

I think it is crucial to underline how the evolution of censorship focused on AI, which makes state control over public discourse even more sophisticated, in particular at a time when models of Chinese AIs such as Deepseek make waves, said Xiao, the researcher of Berkeley, in Techcrunch.