From Sampling to Surveillance: Evaluating the Effectiveness of Artificial Intelligence in Detecting Complex Financial Fraud

Abstract

For most of the twentieth century, the detection of financial fraud rested on an uncomfortable compromise: because no auditor or investigator could examine every transaction, assurance was built on samples, and fraud that fell outside the sample escaped notice. Artificial intelligence promises to dissolve that compromise by subjecting entire populations of transactions, disclosures, and communications to continuous algorithmic scrutiny. This article evaluates how far that promise has been kept. Drawing on three decades of empirical research in accounting, information systems, and computer science, it examines the performance of supervised classifiers, anomaly detection, natural language processing, and network analytics against complex schemes such as financial statement manipulation, collusive procurement fraud, and layered transaction fraud. The evidence supports a qualified conclusion. Machine learning models now outperform traditional ratio-based screens by meaningful margins, yet their effectiveness is constrained by severe class imbalance, biased training labels drawn only from detected fraud, adversarial adaptation by offenders, and opacity that sits awkwardly with evidentiary standards in criminal and regulatory proceedings. The article argues that artificial intelligence is best understood as an instrument of triage rather than adjudication, and it draws out the governance, forensic, and pedagogical consequences of that position for both mature and emerging markets, including African jurisdictions such as Ghana.