On a clear morning in Shanghai’s Pudong district, a camera detects a crowd assembling near a subway exit. Within seconds, an AI system flags the gathering, cross-references faces against a national database, and fires a pre-emptive alert to local police — before a single word has been spoken, let alone a permit requested. This is not a speculative scenario. It’s the operational reality of China’s surveillance architecture today, and it’s being rebuilt from the ground up with generative AI, large language models, and a political mandate to make authoritarian control faster, cheaper, and effectively invisible.

The Surveillance State Finds Its Intelligence Layer

China has spent two decades constructing what is almost certainly the world’s most extensive state surveillance infrastructure. Estimates put the country’s camera count at up to 600 million — roughly three cameras for every seven citizens. But raw hardware counts have never been the story. The real transformation is happening in the software layer.

Beijing’s 15th Five-Year Plan (2026–2030), unveiled at the March 2026 “Two Sessions” legislative meetings, enshrines AI-driven governance as a national strategic priority, carrying an explicit directive for China to seize the “commanding heights of science and technological development.” The plan formalises what researchers had already been documenting for two years: an accelerating fusion of generative AI, large language models, and legacy surveillance hardware into a single, predictive control apparatus.

Crucially, China’s amended Cybersecurity Law — the first major revision since 2017 — took effect in January 2026, weaving AI explicitly into the legal architecture of state surveillance for the first time. The Cyberspace Administration of China described the updated law as providing the foundational framework for “cyber sovereignty,” stressing its role in Xi Jinping’s directive for China to become a cyber superpower. When AI and censorship law merge, the implications don’t stay inside China’s borders for long.

1 — The Core Development: How China’s AI Surveillance Network Is Being Rebuilt

The China AI surveillance network upgrade is not a single programme. It’s a layered modernisation of interconnected systems, each accelerated by the same generation of tools now reshaping industries worldwide.

At its foundation sit two legacy projects. Skynet (天网工程), deployed primarily in urban centres including Beijing, Shenzhen, and Chengdu, operates as a high-precision facial recognition and automated tracking system — state media once claimed it could scan China’s entire population in under a second, though researchers at Georgetown’s Center for Security and Emerging Technology have noted that such claims “ignore glaring technical limitations.” Sharp Eyes (雪亮工程), launched in 2015 by the National Development and Reform Commission, extended surveillance into rural and semi-urban provinces including Hunan, Henan, Sichuan, and Guizhou, setting a target of 100% coverage of public space by 2020. It went further: integrating private household cameras into centralised monitoring platforms, and in some areas giving local residents access to live security footage — a model of what researchers now term “participatory surveillance.”

What has changed is the intelligence layer sitting on top of that hardware. According to a December 2025 report by the Australian Strategic Policy Institute (ASPI) — granted to The Washington Post for exclusive early access — the Chinese Communist Party is “harnessing AI to make its existing systems of control far more efficient and intrusive.” ASPI senior analyst Nathan Attrill stated: “AI lets the CCP monitor more people, more closely, with less effort. In practice, AI has become the backbone of a far more pervasive and predictive form of authoritarian control.”

The hardware supply chain is equally telling. Hikvision and Dahua together supply roughly one-third of the global market for security cameras and digital video recorders, and Hikvision directly implements Sharp Eyes infrastructure in cities including Xi’an. SenseTime, designated an official “AI Champion” by the party-state, provides facial recognition algorithms feeding into centralised police databases. The 206 System, developed by iFlyTek, analyses criminal evidence and recommends sentences to prosecutors. In Anhui province, prosecutors use AI platforms to draft indictments and flag inconsistencies in dossiers — an end-to-end automation of the charging process.

The architecture is converging toward what analysts described in March 2026 as an AI-driven criminal justice pipeline — surveillance that doesn’t merely observe, but actively adjudicates.

2 — The Analytical Layer: Predictive Control and the Logic of Pre-emptive Suppression

How Is China Using AI for Predictive Policing?

China is using AI for predictive policing through a combination of large language models, neighbourhood grid worker data networks, and real-time social media monitoring. Systems process individuals’ personality profiles, emotional states, and exposure to “negative cultural influences” to forecast social unrest before it occurs — a function previously requiring large human intelligence operations, now automated at scale.

The most significant shift is not the hardware. It’s the move from reactive surveillance — watching and recording — to predictive surveillance, which attempts to identify threats before they materialise. In August 2025, Guizhou Normal University filed a patent proposing the use of OpenAI’s GPT models as a “core reasoning tool” in a system designed to predict “social governance incidents” — the official euphemism for protests and collective petitions. The patent draws on inputs including individuals’ “long-term emotional states” and “degree of exposure to negative cultural influences,” without specifying how that last category would be measured. Any functioning implementation would depend entirely on the pre-existing surveillance infrastructure.

The human network feeding these AI systems is itself a revealing detail. Since early 2025, multiple Chinese institutions have developed tools built on reports from “grid workers” (网格员) — typically paid community workers who monitor assigned neighbourhood grids and upload incident reports in real time through a dedicated smartphone app. AI systems aggregate and analyse that granular social data, giving local authorities a dynamic, block-level picture of sentiment and risk. This is Xi’s concept of social governance operationalised through machine learning: citizens enlisted as data collection nodes in a system that processes their reports at a scale no human bureaucracy could sustain.

The picture is more complicated when one considers the incentive structures for domestic AI firms. Alibaba, Baidu, and Tencent are building multimodal large language models that censor and reshape descriptions of politically sensitive content — not because they’re state-owned enterprises, but because commercial operating licences in China effectively require it. Private companies like SenseTime didn’t survive by resisting the surveillance state. They thrived by building it, and in doing so, became too strategically valuable for either side to disentangle.

What this produces is an AI ecosystem in which the line between commercial product and state instrument has effectively dissolved. That’s the structural condition that makes Beijing’s surveillance ambitions sustainable in a way that brute state spending alone never could have achieved.

3 — Implications and Second-Order Effects: The Export Problem

What Does China’s Surveillance Technology Export Mean Globally?

The global consequences are no longer a projection. Companies including Huawei, ZTE, and SenseTime have invested hundreds of millions of dollars in AI-related infrastructure across Asia, Africa, and Latin America, according to research by the Alan Turing Institute’s Centre for Emerging Technology and Security. These projects — ranging from broadband rollouts to city surveillance systems — come bundled with Chinese-built AI solutions, effectively embedding China’s technical standards and governance norms in host countries.

The Digital Silk Road has become the primary vehicle for this diffusion. When a government in Central Asia or sub-Saharan Africa purchases a “safe city” package from Huawei, it frequently receives the same underlying surveillance architecture deployed in Xinjiang, rebranded and repackaged for export. The technology transfer is also a norm transfer: the proposition that government surveillance of this kind is normal, desirable, and technically achievable.

The supply chain problem runs in both directions. A November 2025 congressional report found that American-made semiconductors, cloud computing resources, and AI development tools continued to flow into Chinese surveillance firms despite existing export controls. Representative Raja Krishnamoorthi argued that Washington had “deprioritised human rights protections in its China policy,” and that tightening controls would require coordination with European and Asian allies — because many of the most advanced AI systems and semiconductor manufacturing tools are produced collaboratively across borders. A unilateral American response, the report concluded, will be insufficient.

Inside China, the judicial implications are concrete and accelerating. Oxford University’s Institute of Technology and Justice has documented that China’s Supreme People’s Court declared all courts to be using AI tools in judicial proceedings by the end of 2025, with full AI integration across the justice system targeted for 2030. In Shanghai, an AI platform now recommends whether suspects should be arrested or granted bail. In at least one prison, facial recognition cameras monitored inmates’ expressions, flagging them for intervention if they appeared angry. Surveillance has moved inside the cell.

4 — The Counterargument: The System Is Less Unified Than It Looks

The instinct of outside observers is to imagine China’s surveillance state as a seamlessly coordinated machine, operated from a single console in Zhongnanhai. The operational reality is considerably messier.

Researchers who study the system closely note that China’s surveillance infrastructure is fragmented — a patchwork of overlapping jurisdictions, incompatible data standards, and uneven local implementation. Skynet and Sharp Eyes were rolled out by different agencies at different times; Xinjiang’s Integrated Joint Operations Platform (IJOP) was built largely in isolation from the national infrastructure. Police Cloud systems vary dramatically between provinces. Academic work published in Regulation & Governance in 2024 documented how platformised policing generated massive datasets that frequently couldn’t communicate with one another — information silos in the middle of a supposed information state.

That fragmentation limits actual predictive capability. Beijing wants unified AI surveillance; it has, for now, a collection of partially connected systems generating data that AI tools are only beginning to stitch together. The gap between the Chinese state’s surveillance ambitions and its operational architecture remains measurable — and that gap is precisely where privacy still partially exists.

Some Chinese legal academics have quietly raised accountability concerns, too. The Supreme People’s Court’s push for AI in sentencing has met internal scepticism from judges who ask: who is responsible when an algorithm recommends the wrong outcome? These aren’t dissident voices; they’re institutional concerns from within the apparatus itself.

None of this adds up to a reassuring counter-narrative. The trajectory is clear, the investment is sustained, and the 15th Five-Year Plan provides the political mandate to accelerate integration. But it does mean the gap between Beijing’s stated ambitions for its AI surveillance network and the system’s operational reality remains wider than official statements suggest. Ambition and capability are not the same thing — and in this domain, treating them as identical is its own form of error.

The Architecture Is Now Legal, Not Just Technical

China’s AI surveillance overhaul is, at its core, the industrialisation of authoritarian control — the application of the same machine learning techniques powering medical diagnostics and content recommendation to the problem of population management. The efficiency gains are real. The harms scale with the efficiency.

What makes this moment distinctively consequential is the legal architecture now surrounding it. The amended Cybersecurity Law, the 15th Five-Year Plan’s explicit directives, and the Supreme Court’s AI integration mandates have moved this from a technology project to a formal governance system. The apparatus is being institutionalised, not just expanded. That distinction matters: institutions survive their architects, outlast political cycles, and are far harder to dismantle than experimental programmes.

Whether democratic governments and technology companies can meaningfully slow the proliferation of these tools — across China’s borders, through the supplier relationships that sustain them, and into the legal frameworks of countries that may find them attractive — is among the defining policy questions of the next decade. Export controls, sanctions, and multilateral coordination are all on the table. None has yet proven sufficient.

The cameras don’t blink. The question is whether anyone watching them will.