TOWARD TRUSTWORTHY AI IN INSURANCE FRAUD DETECTION: A GOVERNANCE AND ACCOUNTABILITY FRAMEWORK FOR EXPLAINABLE MODELS

Abstract

Insurance fraud constitutes one of the most persistent and financially consequential challenges confronting the global insurance industry, with annual losses estimated between USD 80–100 billion in the United States alone. While machine learning and deep learning models have demonstrated superior discriminatory power over legacy rule-based detection systems, their deployment in consequential underwriting and claims adjudication contexts raises acute concerns regarding model opacity, algorithmic accountability, and regulatory compliance. This paper introduces TAXIF (Trustworthy and Accountable explainable Insurance Fraud detection), a governance-integrated explainability framework coupling a stacked ensemble classifier with a multi-layer interpretability architecture, a real-time explanation generation module, and an institutional accountability layer encoded with GDPR Article 22, NAIC model law, and IAIS Insurance Core Principle 19 constraints. TAXIF learns the joint behavioral distribution of claimant demographics, policy history, claims pattern sequences, and network relationship signals through a gradient-boosted ensemble augmented with attention-weighted feature attribution. Controlled experiments on a synthetic dataset calibrated to ISO Claim Search and NICB loss statistics demonstrate: AUC-ROC of 0.978 against 0.931 (single XGBoost baseline) and 0.741 (logistic regression); SHAP fidelity score of 0.94; average explanation latency of 11 milliseconds per claim; and 100% regulatory gate adherence across all 1,000 experimental trials. TAXIF demonstrates that explainability and predictive performance are not in fundamental tension when governance constraints are embedded architecturally rather than appended as post-hoc tooling.