Joint Money-Laundering Detection Across Banks: Shown on Synthetic Data, Gated by Law

The problem

Starting point. The anchor paper in AI in Practice found that machine-learning monitoring of transactions runs inside individual banks, that its strongest result is an operator-reported before-and-after comparison without a control group, and that such a model can only rank the alerts its own bank's rule engine has already raised¹. Each bank's model therefore sees only its own customers. This paper adds a different question: what could change if several banks' models, or at least their warning signals, could be combined on the same suspicious networks.

The gap. The Financial Action Task Force (FATF), the international standard-setter for money-laundering controls, wrote in July 2022 that "A single financial institution has only a partial view of transactions", and that private-to-private sharing outside financial groups "is perceived as restricted in many jurisdictions"². If funds were moved through accounts at several banks, each bank would see only one stretch of the route. No dated figure for the share of laundered funds that goes undetected was opened for this paper, so none is given; the problem is described here only through this structural limit.

What already works

A research prototype. In May 2023 the Bank for International Settlements (BIS) Innovation Hub published Project Aurora, which tested money-laundering detection on a synthetic data set of transactions between financial institutions, individuals and businesses, within one country and across borders³. Synthetic means generated for the test, not real customer records. The project page says the approach detected "potentially up to 3 times more" laundering involving complex schemes and cut false positives, the alerts that turn out to be innocent, by up to 80% (project-reported, simulation on synthetic data)⁴. The report concludes that approaches combining data across institutions were more effective than "the current siloed approach", with the best results in a simulated cross-border set-up where data were protected and consolidated and network analysis was used³.

The report is also open about limits. It says the performance of such analysis "is only as good as the breadth of data available", that the centralised simulation does not reflect "data protection, data localisation, legal, regulatory and other factors", and that a federated approach "may offer an alternative solution, however there could be a trade-off with effectiveness"³. Federated learning is a method in which each bank trains a model on its own data and shares only what the model has learned, not the data. In plain terms, the best results came from a set-up that the report itself says ignores legal and data-protection constraints, and the set-up that those constraints favour might perform less well.

Federated learning in payments. Swift, the messaging network banks use for cross-border payments, said in October 2024 that since February it had worked with financial institutions on federated learning combined with privacy-enhancing technologies, and that the resulting fraud-detection use cases were "set to be tested in a sandbox environment" (Swift-reported)⁵. Trade press reported in September 2025 that Swift announced an experiment with 13 banks on ten million synthetic transactions, in which models were "twice as effective" at spotting fraud as models trained on one institution's data (Swift-reported; metric and test design not stated)⁶. These are fraud tests, not money-laundering tests, and neither source describes use on real customer data.

Legal sharing in production. Singapore's COSMIC platform went live on 1 April 2024 with six major banks⁷. It lets banks share information about customers who show multiple "red flags" once stipulated thresholds are met⁷. When the law was introduced in May 2023, the thresholds were described as rising from requesting information, to providing it, to placing a customer on a watchlist, and banks were told not to rely solely on COSMIC information to end a customer relationship⁸. In August 2026 the Monetary Authority of Singapore (MAS) said participating banks had filed additional suspicious transaction reports and closed a significant number of suspicious accounts, without giving figures, and that COSMIC was being widened beyond its three initial risk areas, with the expansion "scheduled for completion within 2 years"⁹. The pages read do not describe a shared machine-learning model on COSMIC.

In the United Kingdom, sections 188 and 189 of the Economic Crime and Corporate Transparency Act 2023 have been in force since 15 January 2024. They are voluntary, and they remove confidentiality and civil-liability barriers to sharing customer information between regulated firms, directly or through a third-party intermediary that may include consortia, but they do not permit anything that would breach UK GDPR¹⁰. In the Netherlands, five banks set up Transaction Monitoring Netherlands (TMNL), where alerts go only to banks in the payment chain (trade press, 5 August 2020)¹¹. The anchor paper records that TMNL's use of machine learning was described as envisaged, not confirmed¹. A law-firm note of 19 August 2021 recorded questions about data privacy, competition and whether handing monitoring to a utility is outsourcing a critical function¹².

The European Union has also written a route into law. Article 75 of Regulation (EU) 2024/1624, the EU anti-money-laundering regulation, provides for partnerships for information sharing. In the unofficial copy read, it limits what can be shared to listed categories, allows sharing only for customers tied to higher-risk situations or to those about whom more information is needed, requires supervisors to verify the arrangements and a data protection impact assessment before activities begin, allows information generated by machine learning to be shared only with adequate human oversight, and bars sole reliance on shared information¹³. The EUR-Lex record of the regulation lists 10 July 2027 as the date of first entry into force¹⁴.

The possibility

One way this could work is a partnership of banks in one jurisdiction, set up under a legal gateway of the kind described above. Each bank would keep its transaction and customer data in its own systems. A model would be trained across the partnership by federated learning, so that the learning travels to the data rather than the data to a central store, and a coordinating body would combine the results. What each bank would receive could be a risk score or flag for accounts and entities that look like part of a network spanning several banks, without any bank receiving another's records. Investigators at each bank would then assess the flag and make the decision about the customer, as COSMIC guidance and the EU regulation require⁸¹³.

The role of AI. The AI would be the model that finds network patterns, the kind of network analysis used in Project Aurora³, and the federated training that lets it learn across banks. Everything else would be ordinary systems, law and money: the legal gateway, a way to recognise the same person or company at different banks without exposing identity, secure infrastructure, audit trails, supervisor sign-off and funding. If most of the gain came from banks seeing each other's red flags, the machine-learning part of the idea could be the smaller half. COSMIC arranges red-flag sharing without a joint model being described⁷, and the parts of the Aurora report read in full do not separate how much of its improvement came from the model and how much from the wider view of the data³.

What would have to be true

1. A legal route to share exists and fits joint model training. The UK measures leave UK GDPR in force and cover UK-based sharing only¹⁰. The EU partnership article limits sharing to listed categories and to higher-risk customers, and requires an impact assessment before any personal data are processed¹³. Whether training a joint model on all of a bank's customers would count as sharing under such provisions would have to be settled; the sources read do not say.
2. Protected or federated methods keep most of the performance. Project Aurora's strongest result came from protected and consolidated data in a simulation that ignored legal limits, and the report says a federated approach could trade away effectiveness³.
3. The results hold on real data. Every performance figure found comes from synthetic data³⁶. A test on real transactions, run alongside existing monitoring, would have to show that the gain survives messier data and real criminal behaviour.
4. Supervisors and data protection authorities agree before the start. Under the EU article, supervisors verify the mechanisms and the impact assessment before the partnership begins¹³.
5. Humans keep the decision. COSMIC says banks should not rely solely on its information to end a customer relationship⁸, the EU article bars sole reliance and requires an own assessment¹³, and the UK guidance says its measures are not designed to give sectors additional powers to exclude customers inappropriately¹⁰.
6. Costs are covered and each bank has a reason to join. The sources read give no cost figures for COSMIC, Aurora-style systems or a federated model, so who would pay is open.

What could stop it

What the evidence shows

• Synthetic data only. The Aurora results (2023) and the Swift figures (reported 2025) rest on synthetic transactions³⁶. No result from a live, shared machine-learning model for money laundering was found in the sources read.
• A stated trade-off. The Aurora report says a federated approach could trade away effectiveness³, and the Swift figure comes without a stated metric or test design (Swift-reported)⁶.
• Thin results from live platforms. More than two years after launch, MAS's August 2026 reply gave no counts for COSMIC and said it must balance information sharing against "undue de-risking and confidentiality of customer information"⁹.
• Contested legality in the Netherlands. A compliance adviser's article of 11 March 2024 reports "strong criticism" of TMNL from the Council of State and the Dutch Data Protection Authority over account holders' privacy, and says a Dutch newspaper reported on 25 February 2024 that TMNL "may not continue in its current set-up" (adviser's account; no court or regulator ruling is cited)¹⁵. A June 2022 statement by the Dutch Banking Association says the Council of State advised against joint transaction monitoring and questioned whether sharing on that scale is necessary (translated from Dutch)¹⁶. TMNL's status after March 2024 was not found.
• Privacy technologies not yet at scale. FATF wrote in 2022 that privacy-enhancing technologies "have not currently been adopted at scale" and that "adequate governance and legal frameworks are key"².
• Limits written into law. The UK measures are voluntary, domestic and do not override UK GDPR¹⁰. The EU partnership article restricts sharing to higher-risk situations and requires human oversight of machine-learning output¹³.
• Searches for failure. Searches for court rulings against bank data-sharing platforms found none in the sources read. That is an absence of findings, not evidence of absence.

Reasoning

If a joint model flagged a customer, the flag could follow that customer from bank to bank. A wrongly flagged person could then lose several accounts at once, which is the kind of undue de-risking MAS says it is balancing against⁹. Those who would bear the risk are customers with little ability to challenge a score they cannot see.

Incentives could also work against the idea. Each bank would carry the cost of joining and the legal risk of sharing, while much of the benefit, such as disrupted networks, could accrue to the wider public and to law enforcement. A bank that already has strong monitoring might gain little from pooling, and smaller banks might gain most but have the least capacity to join. If criminals moved activity to institutions outside the partnership, the pooled view could lose value over time.

Simpler alternatives could do part of the job. If structured red-flag sharing, as in COSMIC and the UK measures, delivered most of the benefit, the case for a joint machine-learning model would shrink to whatever extra detection it added on real data, and the sources read give no measure of that. The strongest case for the idea may therefore be narrow: networks that only a model can see, and that red flags alone would miss.

Signs to watch

• Completion of COSMIC's expansion, which MAS said in August 2026 is scheduled within two years, and any published figures on alerts, reports or accounts⁹.
• Results from Swift's sandbox tests of federated learning, especially any that use real data or address money laundering rather than fraud⁵⁶.
• Outputs from the second phase of Project Aurora, whose project page lists an open call in July and August 2025 and a workshop on privacy-enhancing technologies in October 2025⁴.
• The EU's 10 July 2027 date of first entry into force, as listed by EUR-Lex¹⁴, and the first partnerships notified to supervisors under Article 75¹³.
• Whether the Data (Use and Access) Act 2025 provisions on recognised legitimate interests for crime-related sharing, stated in UK guidance to come into force in 2026, are in force¹⁰.
• Any decision on TMNL's future set-up in the Netherlands¹⁵.