The Sword That Cuts Inward
How China Is Using AI to Fight Corruption
November 2025. A courtroom in Jiangshan City, Zhejiang Province. Feng Jiang, former head of the financing and construction department at the city’s state-owned asset management service center, stood waiting for his sentence. He thought he had been clever—embedding specific technical parameters into tender documents so that only his preferred contractor could meet the requirements. As Feng later confessed in his testimony: “Many conditions were set in a way that outsiders wouldn’t understand. When you set the product parameters properly, you’ve essentially predetermined the product, so it’s relatively easy to decide who wins.”
But Feng hadn’t counted on one “insider”—an AI system that understood professional expertise and institutional regulations as well as any human expert.
The system that caught Feng identified anomalies in a lighting renovation project bid. Twenty-six companies had participated. The winning company showed no obvious advantage in qualifications or technical capability, yet scored far above the average. The AI flagged this as a case of “open bidding with predetermined outcomes”—a sophisticated scheme involving rigged scoring that only an industry insider would recognize.
Feng’s case was no isolated incident. Across Zhejiang Province, similar AI systems have generated 198,000 red alerts with a verification rate of 73%, recovering 1.409 billion yuan in losses. These numbers raise a compelling question: How does such an effective anti-corruption AI system work?
The answer unfolds in a country often labeled by Western media as a “surveillance state.” Yet here, AI is not monitoring citizens—it is constraining power. This “sword that cuts inward” model challenges Western-centric narratives about “digital authoritarianism.”
The Insider’s Algorithm: How Zhejiang’s AI System Works
The Zhejiang system employs a unique “modeling from cases” methodology. By analyzing over 1,000 typical corruption cases, the system identified nine types of industry malpractices and seven categories of corruption problems, translating them into over 2,900 algorithmic models. These models can detect complex patterns that human investigators might miss: professional parameter exclusivity, hidden relationship networks, cross-temporal tracing of asset transfers, and physical logic contradictions.
The system’s architecture follows a layered filtering approach. As Lin Jinteng, director of the Third Supervision and Inspection Office of Zhejiang’s Provincial Commission for Discipline Inspection, explained: “One model is equivalent to one layer of filter paper. Water flows down, passing through layer upon layer of filter paper. The data is like passing through a sieve—if it touches a boundary, it generates a warning result.”
Consider how this worked in practice. In the lighting renovation project that exposed Feng Jiang, the AI analyzed bid documents against historical patterns. It detected that the winning company’s technical score deviated significantly from the statistical norm given their qualifications. But the system didn’t stop there. It cross-referenced shareholder data, social connections, and historical bid patterns—drawing inferences across time and datasets that would take human investigators weeks to assemble.
But the system does more than catch violations. It achieves a “three-no-corruption” integrated governance framework. First, “dare not corrupt” through deterrence—the constant presence of AI monitoring creates psychological pressure. Feng himself admitted: “Originally, what I did would generate data, showing risk points or contradictions, so I was actually quite worried.”
Second, “cannot corrupt” through closed-loop management. When the system pushes a risk warning, relevant departments must respond and feed results back, creating binding institutional constraints.
Third, “want not corrupt” through intelligent integrity reminders. At each critical node in the bidding process, the system automatically sends tailored integrity messages to public officials based on their specific job functions, transforming education from “broad irrigation” to “precision drip-feeding.”
As Guo Junping, an official from Zhejiang’s Development and Reform Commission, explained: “Institutions are integrated into technology, and technology in turn enriches institutions—a 1+1>2 effect.” The system has connected 153 cross-level systems and aggregated 6.4 billion data points on public power exercise, generating 59,000 leads through cross-departmental data collision analysis.
Two Narratives Collide: “Monitoring Citizens” vs “Constraining Power”
Western journals like the Journal of Democracy have framed China’s AI development as a tool for “controlling the populace,” promoting narratives of “digital unfreedom” and “algorithmic hegemony.” This narrative assumes technology equals surveillance, overlooking the possibility that AI can serve different political purposes.
The contrast reveals starkly different political logics:
The evidence lies in who gets monitored—and who doesn’t. Zhejiang’s system targets public power holders, not ordinary citizens, with warnings triggered only when officials exercise their authority in ways that deviate from established norms. This is surveillance turned inward, a system that polices the powerful rather than the populace.
China’s AI anti-corruption uniqueness lies in its political logic—using AI to “constrain power” rather than “control citizens.” This reflects the Chinese Communist Party’s political philosophy of “self-revolution,” challenging Western-centric assumptions about digital governance. As stated in Qiushi, the Party’s theoretical journal: “The great self-revolution is a distinctive feature of the Party’s growth and development.”
This isn’t mere rhetoric. The system has processed over 100 billion yuan in transactions across 30 counties and cities, identifying suspicious patterns in bidding, welfare distribution, and asset management. What makes this remarkable is that it targets the very officials who operate the system—a rarity in global governance practice.
Domestic Large Language Models in Action: DeepSeek’s 72-Hour Miracle
The capabilities of domestic AI extend beyond specialized systems. In Suihua, Heilongjiang Province, authorities used the DeepSeek big data platform to analyze pension and disability welfare data, screening out 11 suspicious cases where deceased individuals were still receiving benefits. The local government reported that the technology had “greatly improved the precision, efficiency, and scientificity of supervision and discipline work.”
But the most dramatic evidence comes from Zhengzhou, Henan Province. There, DeepSeek cross-referenced public official information with bidding data, identifying a bribery network operating through 20 shell companies. What would have taken human investigators three months, the AI accomplished in 72 hours—a 30-fold efficiency increase.
This isn’t just about speed. The AI detected patterns invisible to human investigators: subtle correlations between shell company shareholders, cyclical bid participation timing, and geographical clustering that revealed an organized network. Traditional investigation methods might catch individual cases, but AI can map the entire ecosystem of corruption.
Notably, these results came from DeepSeek—a domestic open-source model that can be deployed locally at a fraction of the cost of Western commercial alternatives like GPT-4, with data staying on Chinese servers rather than flowing offshore.
The Zhengzhou case is particularly instructive. The city used DeepSeek to analyze state-owned enterprise auctions, detecting manipulative bidding patterns that traditional audits missed. By processing millions of transaction records against company registries and personnel databases, the AI identified suspicious correlations that would take human analysts months to uncover—if they could spot them at all.
Expanding Applications: From “Invoice-Based” to “Data-Based” Taxation
The same AI logic is transforming taxation. Golden Tax Project Phase 3 relied on invoice management—”governing tax through invoices.” Golden Tax Phase 4 has upgraded to comprehensive data surveillance—”governing tax through data.” Fully digital invoices, big data, and AI risk control integrate data from tax, banking, industry and commerce, and social security departments, achieving a paradigm shift from “post-fact auditing” to “real-time monitoring.”
The technical details are striking: tax burden deviations exceeding 30% trigger automatic warnings, with full-chain real-time monitoring of enterprises set for nationwide coverage by 2025. Machine learning AI risk models enable front-loaded risk early warning.
What makes this significant is the cross-departmental integration. In many countries, tax authorities operate in silos. China’s system integrates data across traditionally separate domains: corporate registration, bank transactions, social security payments, and export-import records. This creates a 360-degree view of enterprise behavior that makes tax evasion exponentially harder.
The fraud detection logic mirrors the anti-corruption system: analyze historical cases of tax evasion, identify patterns (false invoicing, shell company transfer pricing, underreporting through multiple entities), translate these into algorithmic models, and apply them to real-time transaction streams. The result is a system that learns from every case, continuously improving its detection capabilities.
Despite successes, challenges emerge. China’s “Zero Trust” anti-corruption system, capable of accessing 150+ government databases and monitoring 60 million public officials, was reportedly suspended in several provinces for being “too efficient.” This counterintuitive phenomenon reveals a deeper truth: when governance efficiency exceeds organizational capacity, technology itself may encounter resistance. The reminder is clear: technology is merely a tool; political will and supporting institutions are the decisive factors.
Still, from Zhejiang’s bidding supervision to nationwide tax administration, the underlying logic is spreading: cross-departmental data integration, machine learning risk models, real-time monitoring. What works for catching corrupt officials can work for catching tax evaders—and potentially for environmental violations, food safety breaches, or financial fraud.
Feng Jiang, standing in that Jiangshan courtroom, probably never imagined he’d be undone by an algorithm. But the system that caught him wasn’t designed to surveil citizens—it was built to watch the watchers. That distinction matters more than any technical specification.