A Chinese surveillance technology company has been developing artificial intelligence systems designed not only to monitor political dissidents but to identify who might become one - generating behavioral profiles of citizens and using AI models to flag individuals as potential political risks before they have taken any public action. The work, still apparently in a research phase, represents one of the most explicit known efforts to move authoritarian surveillance from reactive observation to predictive suppression. Documents from Geedge Networks, a company that sells a commercial version of China's Great Firewall, reveal the scope and ambition of this effort.
The Architecture of Anticipatory Control
Geedge's existing commercial products allow governments to monitor internet traffic and detect when users attempt to circumvent censorship - tools that have already been exported to countries including Ethiopia, Kazakhstan, Myanmar, and Pakistan. But the newer research documented in leaked company files goes considerably further. In the first months of 2024, Geedge researchers were working to build behavioral profiles drawing on telecommunications records, social media activity, and location data. AI models were then applied to classify individuals and, in the company's own framing, "detect harmful information" - a phrase the Chinese Communist Party uses as standard terminology for political dissent or content the state wishes to suppress.
The predictive ambition places this work in a category qualitatively different from conventional surveillance. Conventional monitoring records what a person has said or done. Predictive systems attempt to assess what a person is likely to say or do, then surface that person for preemptive scrutiny. "This is what happens when mass surveillance meets AI," said Brett J. Goldstein, director of the Wicked Problems Lab at Vanderbilt's Institute of National Security, whose researchers identified and analyzed the Geedge documents. The dystopian character of such a system is not merely rhetorical: a government acting on predictive scores could detain, restrict, or harass citizens who have committed no act and expressed no public view.
Geedge's research arm, Mesa Lab, conducted this work with government support - an arrangement that mirrors the close integration of Chinese state security priorities with nominally commercial technology firms. A comparable effort by another Chinese company, GoLaxy, was documented separately by Vanderbilt and The New York Times: GoLaxy was developing AI-powered software to push targeted propaganda aligned with Chinese government positions while suppressing opposing viewpoints. Together, these efforts outline an emerging ecosystem of AI-assisted information and population control.
Where Export Controls Have Slowed the Project
The most ambitious version of Geedge's predictive system faces a significant material constraint: computing power. Training and running the kind of large-scale AI models that could process intercepted phone calls, surveillance video, and cross-referenced behavioral data at population scale requires substantial quantities of advanced graphics processing units - the specialized chips that power modern artificial intelligence. US export controls introduced during the Biden administration restricted China's access to the most advanced chips, including those designed by Nvidia.
American officials assessing Geedge's capabilities have concluded that the company has enough processing power for its current surveillance products but would need considerably more advanced chips to execute the full predictive surveillance vision. That gap, for now, appears to be acting as a practical ceiling. The Trump administration has modified some of the Biden-era export restrictions while maintaining controls on the most powerful Nvidia processors. During President Trump's recent trip to Beijing, US officials indicated that China would gain access to a more advanced tier of Nvidia chips - a development that complicates the picture. China is simultaneously investing in domestically designed AI chips with the explicit goal of removing its dependence on US-origin hardware, which would render export controls ineffective as a long-term check on these ambitions.
The Broader Surveillance Export Market
Geedge is not a fringe actor. Its core product - a commercial adaptation of the Great Firewall - has been deployed across multiple countries to conduct mass surveillance on mobile networks. The 100,000 documents leaked in September and subsequently analyzed by Wired and other publications trace a pattern of technology transfer from China to governments with poor human rights records. This is the commercial infrastructure of digital authoritarianism: software built for domestic control, packaged and sold internationally.
China's Public Security Bureaus have also been moving to integrate DeepSeek, the Chinese AI model, into predictive policing applications - a parallel development that suggests the Geedge research is part of a wider institutional push rather than an isolated experiment. The convergence of mass data collection, increasingly capable AI, and a state apparatus willing to deploy both against its own population presents a challenge that export controls alone cannot resolve. Technology, once developed, transfers. Architectural knowledge spreads. And governments that acquire these tools rarely relinquish them.
For civil liberties advocates, the scenario Geedge is working toward represents a fundamental shift in the relationship between the state and the individual - one in which the threshold for state action against a citizen is no longer what that person has done, but what an algorithm estimates they might do. That is not a technical refinement of existing surveillance. It is a different kind of power.
