AI Researchers Are Training Machines to Predict How Money Launderers May Move Next

Artificial intelligence is becoming a serious part of the anti-money laundering conversation, but the newest development is not simply about making transaction monitoring faster.

The more important shift is that researchers are now trying to train machines to anticipate how money launderers may adapt before those methods appear in real financial data.

A research team at the University of Southern California’s Viterbi School of Engineering is working on a project funded by DARPA that aims to generate fictional but realistic money-laundering schemes. The purpose is not to describe past criminal behaviour, but to create plausible future laundering scenarios that can be used to test and improve detection systems.

That distinction matters. Most AML systems are still built around known patterns of suspicion. A bank or regulated business may look for unusual transaction volumes, payments below reporting thresholds, rapid fund movements, high-risk jurisdictions, shell companies, circular transfers, unexplained third-party payments or activity that does not match the customer’s expected profile. These are useful indicators, but they are mostly reactive. They depend on patterns that have already been identified.

Criminal networks, however, do not remain static. Once a typology becomes widely known, sophisticated actors can adjust. They change payment channels, alter transaction size, use different intermediaries, move into new sectors or exploit gaps between institutions and jurisdictions. This creates a constant delay between the evolution of laundering behaviour and the controls designed to detect it.

The USC project is aimed at reducing that delay. By building systems that can generate realistic laundering scenarios, researchers can stress-test AML models against behaviour that may not yet exist in the historical data available to financial institutions. In simple terms, the idea is to make detection systems train against imagined criminal strategies before those strategies become real cases.

This approach connects with DARPA’s wider interest in anticipatory and adaptive anti-money laundering technology. The agency has described the current AML environment as too manual, too reactive and too expensive. Its programme is focused on replacing slow analytical processes with more agile, algorithmic methods capable of identifying complex laundering activity at scale.

The reason this is necessary is that money laundering is rarely a single suspicious transaction. It is usually a process. Funds may move through a sequence of accounts, companies, digital wallets, payment processors, professional intermediaries and commercial contracts. Each individual step may appear ordinary. The suspicious element often appears only when the entire network is examined together.

That is why AI and machine learning are attractive to AML researchers. A human analyst may review alerts one by one, but a machine can analyse broader patterns across time, entities and relationships. It can identify connections that are not obvious in isolation, such as repeated counterparties, unusual timing, mirrored transaction flows, shared addresses, connected directors, wallet clustering or transaction behaviour that changes after a compliance trigger.

Synthetic data is becoming increasingly important in this field. Financial institutions cannot easily share real customer transaction data because of privacy, confidentiality, banking secrecy and regulatory restrictions. Even where data can be anonymised, the risk of exposing sensitive information remains high. This limits the ability of researchers and technology providers to train and benchmark AML models using realistic data.

Synthetic AML datasets attempt to solve that problem by creating artificial transaction data that resembles real-world financial behaviour without exposing real customers. These datasets can include legitimate activity, suspicious activity, network structures, timing patterns and laundering scenarios. They allow researchers to test whether detection models can identify illicit patterns while avoiding the legal and privacy problems linked to real bank data.

However, synthetic data is only useful if it is realistic. If the generated laundering schemes are too simple, the model may perform well in testing but fail in the real world. Real laundering activity involves partial information, incomplete labels, changing behaviour, hidden beneficial ownership and transactions that may be deliberately designed to look normal. A useful synthetic model must capture that complexity.

This is where the new research direction becomes more interesting. Instead of merely creating artificial data, researchers are trying to create adversarial behaviour. The machine is asked to think like a launderer: how would someone hide funds, avoid detection, move value, create distance from the original source and integrate money into the legitimate economy? Once those fictional schemes are created, AML systems can be tested against them.

For compliance teams, the practical benefit could be significant. A regulated firm could use this type of technology to test whether its transaction-monitoring rules are too narrow, whether its risk models miss network-level behaviour, or whether its controls depend too heavily on old typologies. This could improve not only detection but also internal audit, model validation and regulatory defensibility.

At the same time, AI is not a magic solution to financial crime. A system that produces alerts without explainable reasoning can create new problems. Compliance officers still need to understand why a transaction, customer or network was flagged. Regulators will expect a firm to explain its decisions, document its risk assessment and show that its model is governed properly.

There is also the risk of false positives. AML teams already face large volumes of alerts that require manual review. If AI simply produces more alerts without improving their quality, it may increase the burden on compliance departments instead of reducing it. The real test is not whether a model can identify theoretical suspicion, but whether it can help firms prioritise meaningful risk.

The development also raises a governance issue. If machines are used to generate laundering scenarios, firms must ensure that the outputs are controlled, used defensively and not exposed in a way that could assist bad actors. The same technology that helps banks prepare for future laundering methods must be handled carefully so that it does not become a blueprint for evasion.

The wider trend is clear. AML is moving from static rules toward adaptive testing. Regulators and institutions are no longer asking only whether a system can detect known red flags. They are increasingly asking whether it can detect complex behaviour, learn from new patterns and remain effective against criminals who adapt.

For banks, payment institutions, crypto businesses, gambling operators and professional service providers, this shift is important. Criminals already exploit the gaps between sectors. Funds can move from a bank account to a crypto wallet, then to an offshore company, then into gambling activity or real estate. Detection systems that examine each sector separately may miss the overall laundering chain.

AI-based AML research is therefore not only a technology story. It is a response to the reality that financial crime has become networked, cross-border and highly adaptive. The next generation of AML tools will need to understand not just individual transactions, but the behaviour and structure behind them.

The key message is simple: the fight against money laundering is becoming more anticipatory.

Instead of waiting for criminals to reveal the next typology, researchers are trying to model it first. If successful, that could change how financial institutions prepare for laundering risk and how regulators evaluate the strength of AML systems.

By fLEXI tEAM