In China, evidence has emerged that researchers sought to develop artificial intelligence (AI) technology to screen in advance people who might criticize the government or join anti-government protests, like in the film "Minority Report." It is a surveillance infrastructure that scrapes citizens' past and present traces—such as call records, internet browsing history, social media activity, and location data—to calculate the probability that someone will become an opponent.
The New York Times (NYT) reported on the 1st (local time) that, based on more than 100,000 leaked documents from the Chinese surveillance company Geedge Networks and analysis by a Vanderbilt University research team, the company had, since early 2024, been researching technology with the government-backed research organization MESA Lab to identify politically risky individuals with AI.
Geedge Networks was a company that sold a commercial version of China's national internet censorship system, the Great Firewall. It had mainly sold software that blocks access to certain sites or fundamentally prevents the use of virtual private networks (VPNs) that circumvent censorship. The chief scientist at Geedge Networks is Professor Fang Binxing, known as the "father of China's internet censorship." InterSecLab, a security research institute that analyzed the leaked documents, said the company supplies systems equipped with deep packet inspection (DPI), which examines the contents of user data packets, and real-time monitoring of mobile subscribers.
With data collected through this surveillance net, Geedge Networks estimated who could become a political risk in the future. The system the company developed ties together data scattered across digital spaces to grasp user behavior flows. It first links a real-name mobile phone number to a messenger account. It then adds the device's unique identifier and IP address to merge even the history of VPN circumvention into a person's profile. When base station records and GPS signals are added, a time-series trajectory is completed that shows when an individual passed by a protest site and which activist they had contact with. Experts said that in an environment like China's, where the state monopolizes telecom networks and compels platform cooperation, such merging is easy.
If behavior such as frequent VPN use or mainly viewing foreign news sites is combined with repeatedly visiting past protest locations or having contacts with dissident activists, a "potential political risk individual" score is assigned. According to this system, even what books a person bought and which theater they saw which movie in can be identified by tying it to their route. Brett Benson, a political science professor at Vanderbilt University, said, "Seemingly ordinary data becomes the material that determines who the user is and what they will do next."
According to the leaked documents, the researchers discussed ways to identify people's "intentions" and find "harmful information" at a meeting on Feb. 5, 2024. In the Chinese Communist Party system, harmful information goes beyond serious criminal information to encompass political dissent, criticism of the system, and sensitive controversies the authorities seek to suppress. Jimmy Goodrich, a senior research scholar at the University of California's Empirical Studies of Conflict Project, said, "China's security organs are experiencing data overload," adding, "The real value of AI lies in sifting out the core threats from the deluge of data." The implication is that since 1.4 billion citizens cannot be directly monitored by humans, AI would first narrow down the targets to examine.
Experts warned that if AI-assigned risk scores translate directly into criminal suspicion, even citizens who have not actually taken part in protests could end up on a blacklist due to algorithmic error. The European Union (EU), to prevent such situations, explicitly bans through its AI Act systems that predict or assess criminal risk based solely on personal tendencies or traits. Brett Goldstein, director of the Wicked Problems Lab at Vanderbilt University, said, "This is what happens when mass surveillance meets AI," adding, "What China is doing to its own citizens now is a preview of the future that could unfold in any country that deploys unregulated AI."
It is not known whether China has completed this technology and deployed it in the field. Experts predicted that China likely has not secured high-performance semiconductors sufficient to run AI prediction models at this scale. As before, filtering banned words from text through internet censorship is possible with limited computing power. But to predict future behavior by tying together phone call wiretap audio, surveillance video, and location data for the entire population requires a GPU infrastructure with massive computing capacity. Vanderbilt University, based on public documents, said, "There is a record that Geedge Networks faced GPU constraints and began using older, lower-performance AI models and chips." This bolsters the assessment that semiconductor export controls to China imposed by the previous Biden administration have slowed the pace of the country's surveillance-state advancement.
Even after President Trump took office, China still has not been able to introduce top-performance Nvidia chips due to U.S. sanctions. Citing U.S. officials, the NYT reported, "While Geedge Networks has secured GPUs for current products, its most ambitious prediction systems will likely require high-performance chips that China finds difficult to obtain."