A Chinese technology company has been working to build artificial intelligence systems capable of identifying political dissidents - not after they act, but before they do. The research, attributed to Geedge Networks and its government-affiliated research arm Mesa Lab, draws on telecommunications records, social media activity, and location data to generate behavioral profiles of citizens and flag those deemed a potential political risk. The implications extend well beyond China's borders: Geedge already sells a commercial version of the Great Firewall to governments across Africa, Central Asia, and Southeast Asia.
Surveillance as a Product - and an Export
Geedge Networks is not a fringe actor. The company markets network security and internet monitoring software to foreign governments, and documents reviewed by Wired and other publications confirm it has supplied surveillance tools to Ethiopia, Kazakhstan, Myanmar, and Pakistan. Those tools allow client governments to monitor internet traffic at scale and detect when users attempt to circumvent censorship - the same basic function China's Great Firewall performs domestically, now available as a commercial service to any willing buyer.
What the newly examined documents reveal is a more ambitious layer of that infrastructure. In the early months of 2024, according to the leaked Geedge files, researchers at Mesa Lab were working to apply AI classification models to large datasets drawn from mobile networks, social platforms, and physical location tracking. The stated purpose was to "detect harmful information" - a phrase the Chinese Communist Party routinely uses as shorthand for political dissent, criticism of the government, or content the state wishes suppressed. The AI systems, as described in the documents, were designed to generate risk profiles of individual citizens rather than simply flag prohibited content after the fact.
Brett J. Goldstein, director of the Wicked Problems Lab at Vanderbilt's Institute of National Security, described the convergence plainly: "This is what happens when mass surveillance meets AI." His team, working with the leaked document archive, identified the scope of what Geedge and Mesa Lab were attempting to build.
The Chip Constraint - and Its Limits
There is one significant brake on Geedge's ambitions, at least for now: computing power. The most demanding version of the company's predictive surveillance system - one that would process intercepted phone calls, surveillance video feeds, and real-time location data simultaneously - would require advanced graphics processing units far beyond what China can currently acquire through legitimate channels. Biden-era export controls on US-designed chips, particularly Nvidia's most powerful processors, appear to have slowed the development timeline meaningfully. American officials briefed on the matter confirmed that Geedge currently has enough processing capacity for its existing products but would need substantially more advanced hardware to realize the full scope of its predictive ambitions.
The Trump administration has since relaxed some of those export restrictions while maintaining controls on Nvidia's highest-tier chips. During President Trump's recent visit to Beijing, US officials indicated that China would gain access to a more advanced class of Nvidia processors than previously permitted - a concession whose implications for AI-driven surveillance technology remain unclear but are not trivial. China, meanwhile, is actively working to reduce its dependence on US-designed chips entirely, an effort that, if successful, would render export controls a diminishing constraint over time.
A Pattern, Not an Outlier
Geedge is not operating in isolation. Vanderbilt researchers and The New York Times previously documented the work of another Chinese firm, GoLaxy, which developed AI-powered software designed to push targeted propaganda aligned with Chinese government positions while suppressing opposing viewpoints. Where GoLaxy focused on shaping information environments, Geedge's research points toward profiling and pre-emptive identification - a complementary but distinct capability that, taken together, suggests a broader ecosystem of AI-assisted political control being developed within China's tech sector.
China's Public Security Bureaus have also been reported to be incorporating the AI model DeepSeek into predictive policing projects, according to experts and government officials. The direction of travel is consistent: artificial intelligence applied not to criminal investigation after the fact, but to anticipating behavior before it manifests. That shift - from reactive surveillance to predictive suppression - is what makes the Geedge case particularly significant. Governments that have historically monitored known dissidents now have access to tools that aspire to surface unknown ones, based purely on patterns of behavior extracted from everyday digital life.
What Comes Next
The technology described in the Geedge documents is still in its research phase, and the computing constraints are real. But research stages end. The architecture being built now - the integration of telecom data, social media signals, and location tracking into AI-powered risk classification - does not require science-fiction hardware to function at a basic level. It requires more advanced hardware only to scale or to add richer data streams such as video or audio intercepts.
For the governments already using Geedge's existing surveillance software, that distinction may matter less than it appears. The behavioral profiling component, if delivered even in a limited form, would represent a qualitative change in how authoritarian states operate: shifting the target of suppression from people who have dissented to people whose patterns suggest they might. That is a threshold that, once crossed, fundamentally alters the relationship between a state and its citizens - and it is one that export controls alone, however carefully maintained, may not be sufficient to hold back indefinitely.
