AI-Driven Tax Fraud Detection and Tax Compliance Analytics: A Review of Methods, Challenges, and Responsible Deployment (Journal of Computer Science and Technology Studies)

Summary

• Replacing the least promising 10% of manual tax audits in Italy with taxpayers selected by an algorithm raised detected tax evasion by up to 39%.p.2
• Algorithmic audit selection in Italy increased the tax evasion amount actually paid back by 29%.p.2
• An IRS audit selection model that did not explicitly use race still audited Black taxpayers at 2.9 to 4.7 times the rate of non-Black taxpayers.p.5
• Tax administrations leverage artificial intelligence and machine learning to flag fraudulent refund claims and identity-related fraud before financial losses occur.p.2
• Unsupervised learning and autoencoders act as a safety net by detecting anomalous financial reporting and new tax evasion tactics without requiring labeled data.p.2
• Responsible AI deployment in public fiscal agencies requires explainable models, fairness testing throughout development, privacy protections, and human oversight.p.1

Key insights

1. Random Forests have consistently performed best among standard supervised classifiers in identifying non-compliant taxpayers.p.2
2. Replacing the bottom 10% of audits with machine-learning-selected taxpayers in Italy improved detected tax evasion by up to 39%.p.2
3. Unsupervised clustering algorithms can uncover novel tax evasion strategies by flagging statistical outliers in unlabeled tax filings.p.2
4. Multimodal graph neural networks capture relational dependencies across entities to detect complex cross-border tax fraud.p.2
5. Opaque risk scoring models risk repeating historical biases, as demonstrated by the Dutch childcare benefits case where non-Dutch nationality raised risk scores.p.5
6. IRS audit selection algorithms that excluded race variables still generated racial disparities, auditing Black taxpayers at 2.9 to 4.7 times the rate of non-Black taxpayers.p.5
7. Zero-knowledge transformer inference allows tax authorities to verify machine learning outputs without exposing sensitive taxpayer information.p.4
8. Human-in-the-loop oversight is critical to ensure automated tax audits remain legally contestable and subject to democratic accountability.p.5