ARTIFICIAL INTELLIGENCE IN FINANCIAL STATEMENT FRAUD DETECTION: A CRITICAL SYNTHESIS OF EMERGING TRENDS, METHODOLOGIES, AND FUTURE DIRECTIONS FOR U.S. PUBLICLY TRADED COMPANIES

Abstract

Financial statement fraud causes systemic damage to U.S. capital markets, investor confidence, and the credibility of public company disclosures. Artificial intelligence (AI) offers promising new detection capabilities, yet its evidence base and operational readiness remain unevenly developed across the literature. This structured narrative review critically synthesizes peer-reviewed articles published between 2020 and 2026 on AI-based fraud detection in SEC-reporting publicly traded U.S. companies, with focused attention on methodological rigor, strength of evidence, and deployment readiness. The thematic synthesis shows five emerging phenomena: (1) ensemble and deep learning models significantly outperform classical ratio screens in structured data; (2) NLP models based on transformer architectures identify fraud signals from narrative disclosures, which are independent of the structured data; (3) multimodal fusion architectures result in the highest detection performance (AUC 0.86–0.91); (4) AI is starting to transform audit risk assessment and regulatory surveillance; and (5) operational deployment is still hindered by challenges with explainability, temporal validation, label quality, and adversarial robustness. Persistent gaps remain in model explainability, temporal validation, adversarial robustness, and regulatory deployment protocols. The NIST AI Risk Management Framework is proposed as an organizing governance structure. Explainable AI, federated learning, and blockchain-based audit trails are identified as priority research investments, with targeted recommendations offered for regulators, auditors, and capital market participants.