The Bank for International Settlements (BIS) this month published the article The use of artificial intelligence for policy purposes on a report (pdf) submitted to the G20 Finance Ministers and Central Bank Governors.
In the paragraph ‘Oversight of payment systems‘ [*] the authors explain that the current system of detecting financial crime (anti-money laundering, ‘AML’, and countering the financing of terrorism, ‘CFT’) is highly inadequate:

While rule-based monitoring is straightforward, it suffers from a number of shortcomings. It is notorious for false positives – vast numbers of flagged transactions that upon investigation turn out to be innocuous. It is not uncommon for 95% or more of anti-money laundering (AML) alerts to be false positives, creating huge workloads for compliance teams (and the financial intelligence units that they file reports to) and likely distracting from real criminal schemes. (…)
Partly due to these reasons, the result is an ineffective, fragmented global payments system. Criminals exploit the fragmentation and complexity of payment networks to hide illicit funds, while banks expend enormous resources on compliance.

It shows that the concepts developed by the FATF and the EU have led to large-scale waste of money.

The solution is to treat every account holder as a potential criminal and to and to drag a dragnet over all financial transactions of account holders at various banks. This concept was tested in the BIS project Aurora:

Project Aurora was launched by the BIS Innovation Hub as a proof of concept to test new techniques for AML monitoring across institutions and borders (BISIH (2023)). (…)
By having a more complete view of transaction networks, AI models can uncover complex money laundering patterns that would evade isolated checks at one bank. (…)

Aurora successfully performed a simulated collaborative analysis. Multiple institutions’ transaction data were combined in encrypted form, and graph neural networks (a type of AI suited for network data) were applied to identify suspicious clusters of transactions. The results are striking. Aurora’s approach proved far more effective than traditional rule-based AML monitoring approaches. According to the published findings, it detected up to three times more money laundering cases involving complex schemes and at the same time reduced false positives by up to 80%.

The reasons underlying the success are instructive and point to the value of combining behavioural analytics and data-sharing. First, instead of focusing on individual transactions, the focus was on the behaviour of networks of transactions (pattern of flows and relationships between senders/receivers) to spot anomalies indicative of laundering rings. This network-centric, behavioural approach improves detection as launderers might hide individual transactions, but their overall network behaviour (eg funds circling through intermediaries and back to the origin, and multiple accounts funnelling to one exit point) can give them away. Second, by pooling data from many institutions (with appropriate privacy safeguards), the ML algorithms have the “big picture” – they can see the crossinstitution connections that a single bank’s system would miss. For example, Bank A might see funds going to Bank B, and Bank B sees them going to Bank C. But neither alone sees A→B→C as one chain. The combined view does, and for some ML techniques like artificial neural networks and graph neural networks, such a holistic view can substantially improve performance.

The authors of the report believe that fundamental rights are respected in the financial dragnet (financial surveillance) that was tested in the Aurora project. The question is whether this is really the case.

[*] Paragraph 3.3, page 7 and further of the report (pdf).