China police-university team claims AI+LLM tool flags crypto laundering with ~90% accuracy
The research lands as prosecutors say 3,259 people faced 2025 cases tied to virtual-asset laundering and underground finance.
Researchers at China’s National Police University say they built an AI system that uses large language models to flag suspected crypto money laundering with “roughly 90% accuracy.” The claim arrives alongside a 2025 prosecution tally that suggests virtual-asset laundering cases are already being pursued at scale despite China’s trading and mining bans.
China Police-University Researchers Claim ~90% Accurate AI+LLM Crypto Laundering Flags
Researchers at China’s National Police University, an institution operating under the Ministry of Public Security, are described as developing a blockchain-forensics detection framework aimed at identifying money laundering and other economic crimes involving virtual assets such as Bitcoin. The headline metric attached to the work is a reported hit rate: “the system achieves roughly 90% accuracy in flagging suspicious transactions.”
That accuracy figure is presented as coming via a South China Morning Post report, but the packet contains no methodology, dataset description, or definition of what “accuracy” means in practice. The same write-up also flags the operational caveat that matters for traders and compliance desks: a model that is “90% accurate” can still generate false positives, and law enforcement still has to validate algorithmic flags through manual review and legal process.
China’s enforcement backdrop is not theoretical. China’s Supreme People’s Procuratorate reported that in 2025 alone, “3,259 people were prosecuted on suspicion of money laundering involving virtual assets and underground finance.” Even if the new framework is only a lead-generation tool, that prosecution count implies there is enough case volume for automation to be operationally relevant if it gets adopted.
How the AI+LLM Framework Is Supposed to Spot Layering and Structuring
Mechanically, the framework is described as combining AI-driven pattern recognition with large language model (LLM) capabilities to analyze transaction flows across blockchain networks. The AI component is positioned as the pattern engine, scanning large volumes of on-chain data for behavioral anomalies that rule-based monitoring can miss.
The targets are classic laundering behaviors. “Layering” is the tactic of moving funds through many hops, wallets, and transactions to make the origin harder to trace. “Structuring” is breaking activity into smaller pieces or odd patterns to avoid thresholds and obscure intent. The system is described as looking for anomalies consistent with those behaviors across transaction graphs, rather than relying only on static rules.
The LLM piece is framed as the context layer. A large language model is a text-trained model that can interpret and generate language, and here it is described as helping interpret unstructured information tied to transaction activity, such as transaction memos or communication patterns, to help investigators build a narrative around flows. That matters because enforcement outcomes are not scored on how many addresses get flagged. They are scored on whether a case can be assembled, attributed, and prosecuted.
The write-up also states cryptocurrency trading and mining are banned in China, while holding digital assets is not criminalized. The posture described is a split: restrict public market access, but keep building investigative capacity around virtual-asset flows.
What Traders Should Monitor as China’s On-Chain Surveillance Tooling Advances
The first signal that changes the risk profile is deployment. The packet does not say whether this framework is research-stage, piloted, or already used in live investigations, and it does not name which investigative units would run it. Any disclosure that it is operational, and where, would matter more than another accuracy headline.
The second signal is technical specificity behind the “roughly 90%” claim. Traders should look for a primary link to the underlying reporting or any release that defines the dataset, the labeling standard for “suspicious,” and the false-positive and false-negative rates. Without that, the number is directionally interesting but not a performance guarantee.
The third signal is enforcement intensity. The 2025 prosecution total provides a baseline for scale, so the next annual tally or major case announcements that explicitly reference on-chain analytics would help confirm whether tooling like this is being used to prioritize leads.
The fourth signal is downstream compliance behavior. If exchanges and counterparties start tightening around China-linked exposure, it tends to show up as enhanced due diligence notices, account freezes, or new risk rules that cite AI-driven tracing and higher flagging rates.
My Take: Enforcement-Tech Is Rising Faster Than the Public Evidence for This Model’s ‘Accuracy’
The part that matters here is not the “90%” headline, it is the institutional direction of travel. A police-university team under the Ministry of Public Security building AI+LLM tooling for transaction tracing fits a world where China keeps public trading constrained while still treating virtual-asset flows as an investigative surface, and the 3,259-prosecution figure suggests there is already enough throughput for automation to be useful.
The threshold that matters is disclosure: if this framework is confirmed as deployed in live casework and the accuracy claim is backed by a defined dataset and error rates, the compliance risk around China-linked flows stops being narrative and starts being operational. Until then, the practical impact is a higher probability of flags and friction for counterparties that touch China-linked exposure, not a proven step-change in detection quality.
