Artificial Intelligence in Combating Illicit Finance: A Comprehensive Review of TBML, CFT, CPF, and International Sanctions
DOI:
https://doi.org/10.63125/ppkqpy73Keywords:
Artificial intelligence, Machine learning, Illicit finance, Trade-based money laundering, Terrorist financing, Proliferation financing, Sanctions screening, Graph learning, Explainability, Financial Action Task ForceAbstract
Artificial intelligence (AI) is promoted as the answer to a hard detection problem: finding the few unlawful transactions among billions of lawful ones. This review examines its use against trade-based money laundering (TBML), terrorist financing and proliferation financing, and in the application of international sanctions. It is a narrative review of 100 journal articles, whose bibliographic details were verified in Crossref, and of the official instruments. The study first sets out the three threats, the international framework and the rationale for sanctions. The study then reviews seven families of methods: rules, supervised learning, anomaly detection, graph learning, natural language processing, privacy-preserving learning and generative models. The evidence is uneven. Studies on real bank data report that learned models rank suspicious activity better than rule-based monitoring, and studies on a public Bitcoin benchmark report gains over earlier models. For TBML, statistical screening of prices and mirror trade data is established, while machine learning has addressed customs fraud more than laundering. For terrorist financing, proliferation financing and sanctions screening, peer-reviewed evaluation is thin, and in name screening gains in true matches have come with more false ones. Almost no study reports outcomes after deployment. AI research is therefore strongest where data are abundant and weakest where the concern behind sanctions is greatest, a mismatch explained by fragmented data and scarce confirmed cases more than by algorithms. The sanctions literature adds that effectiveness is conditional, evasion is adaptive, and humanitarian harm, de-risking, weak due process and the disputed legitimacy of unilateral measures are documented; we present opposing positions without endorsing either. AI may ease or aggravate these problems, and brings risks of opacity, bias, automated over-compliance and intrusion on privacy. The study offers nine propositions, including investment in linked trade data, explicit costs for false positives, explainable and contestable models, and evaluation of outcomes. Summaries of most cited studies were not checked against full texts, and no new data are reported.


