BEIJING — A Chinese firm long tied to the country’s vast internet controls has set its sights on something more ambitious. Geedge Networks wants to build systems that do not simply watch citizens. They aim to forecast who might challenge the state.
Leaked internal records show the company, seller of a commercial version of the Great Firewall, has explored artificial intelligence tools. These would scan location data, telecom records, social media patterns and online behavior. The goal? Spot signals of future criticism or unrest. The New York Times reported the effort on Monday, drawing on research from Vanderbilt University that examined roughly 100,000 leaked documents.
Researchers described the project as still in development. Yet its direction stands out. Instead of reacting to protests or posts, the technology would act first. It would flag individuals as political risks ahead of any public move. Short. Direct. And, for rights groups, deeply unsettling.
Geedge has supplied tools that help governments filter web traffic and detect attempts to bypass censorship. Its new focus shifts from blocking access to anticipating thought. Meeting notes and progress reports in the leak reveal teams testing large language models. These models would synthesize massive intercepted data sets to build behavioral profiles. One goal: identify intent and uncover what employees called “harmful information.”
From Monitoring to Forecasting
Progress has not come easy. U.S. export controls on advanced chips have forced Geedge to rely on older hardware. That constraint has slowed model training and limited predictive accuracy. The firm runs less powerful systems as a result. Even so, the leaked files indicate persistent effort across research teams. They discussed cross-referencing cell data, movement history and digital footprints to generate risk scores.
Similar ideas appear across Chinese institutions. A December 2025 document from the Fujian Police Academy outlined an AI approach to detect “potential mass incidents” — official language for protests, strikes or gatherings. The system pulls from sound sensors, street cameras, satellite feeds and government reports. It flags buildup early. If it misses something, the AI reviews footage afterward to sharpen future performance. China Media Project examined the filing and related patents filed over the past two years.
Patents from state bodies, universities and private firms describe feeding surveillance streams — social media posts, grid worker reports, noise levels — into models for predictive policing. Huawei has worked on technology that pinpoints locations in photos and builds 3D neighborhood maps. Provincial units in Jiangxi have tested smart terminals that forecast incidents. These efforts tie into the revived Fengqiao Experience, an older model of grassroots control now supercharged with data.
But. The human element lingers. Unpredictable behavior among 1.4 billion people remains hard to model perfectly. Algorithms risk overreach. They can lump petitioners, the unemployed or those with certain social media patterns into high-risk categories. Vulnerable groups pay the price first.
Broader work from the Australian Strategic Policy Institute shows how far the state has already pushed. Its December 2025 report documented AI throughout the criminal justice system. Large language models now handle much online censorship. They scan, flag and delete content in seconds. Human reviewers still step in, but job postings suggest the balance is shifting toward automation. ASPI also highlighted state-backed models for sentiment analysis in Uyghur, Tibetan and other minority languages. The stated aim: better monitor communications at home and among diaspora communities.
Companies such as Tencent, Baidu and ByteDance have developed internal AI censorship tools. Some sell them to smaller platforms. The regulatory environment creates steady demand. And the technology travels. China has exported surveillance systems to authoritarian partners and developing nations. Myanmar offers a documented case. Geedge’s earlier cell-ID tools helped authorities there identify and target pro-democracy figures, according to the Vanderbilt analysis of the leaks.
So the pattern extends. What starts as domestic research can reach clients seeking the same preemptive edge. Brett V., a researcher involved in reviewing the Geedge materials, noted the shift in focus. “They were trying to predict what citizens might do next and with whom,” he said in coverage of the leak.
Western governments watch closely. Export curbs have bought time. They have not stopped the underlying push. Chinese leaders see AI as central to social governance and national security. Premier Li Qiang’s 2024 “AI+” initiative expanded the technology’s role in daily administration. Local experiments multiply. Universities test models that score personality traits for risk. Private firms experiment with open-source models adapted for negative sentiment detection.
Yet accuracy questions remain. Early systems depend on correlation. Frequent visits to certain neighborhoods. Spikes in encrypted messaging. Posts that echo historical critiques. These become inputs. Outputs are risk scores that police or local officials act on. False positives carry consequences. So do misses.
The Vanderbilt team’s work, made public through the New York Times article, adds concrete evidence to long-standing warnings. Earlier ASPI analysis had traced procurement records and corporate filings showing heavy investment in predictive tools. Today’s leaks fill in operational detail. They show engineers grappling with data volume, model cost and hardware limits.
One internal note highlighted the expense. Predictive systems demand enormous computing power. Geedge’s teams weighed trade-offs. Scale back ambition or accept slower rollout. U.S. policy has tilted that calculation.
Still, the trajectory holds. China’s surveillance network already counts hundreds of millions of cameras. Many now feed AI systems. Add location pings, financial traces, health records and social connections. The data pool grows daily. Models improve with volume. And volume China has.
Critics argue this creates a chilling effect. Citizens self-censor not only what they say but where they go, whom they meet, what they search. The state no longer needs to wait for open dissent. It can move at the first statistical whisper.
Recent coverage echoes the concern. NDTV Profit and Moneycontrol both summarized the Vanderbilt findings hours after the Times story broke, noting potential expansion to other governments. No major technical breakthroughs were reported in those pieces. They stressed the preemptive nature and the chip-related slowdown.
Industry voices differ. Some Chinese engineers describe the work as public safety innovation. Others, speaking anonymously in Western reporting, call the predictive leap a nightmare. The gap between those views will not close soon.
What matters now is execution. Can Geedge and its peers overcome hardware hurdles? Will local experiments scale nationally? And how quickly will client states adopt the exported versions?
The answers will shape governance models far beyond China’s borders. Authoritarian regimes have long sought total information awareness. Artificial intelligence offers a path to act on that awareness before threats materialize. Geedge’s research, however imperfect today, sketches one version of that future.
Shortcomings exist. Data bias. Model drift. Human override. These persist. Yet the ambition is clear. Predict first. Intervene early. Maintain control. That logic drives investment, patents and procurement across police academies, tech labs and provincial governments. The leaks provide a rare window into the mechanics. They show a state determined to stay ahead of its people. One data point at a time.
WebProNews is an iEntry Publication