On the economic effects of AI-powered collection systems: the Italian case

This paper examines the economic risks and distributional consequences of implementing Artificial Intelligence (AI) systems in public tax collection, focusing on the Italian case. While AI-powered solutions offer significant potential for improving administrative efficiency and revenue recovery, they introduce substantial challenges regarding algorithmic fairness, transparency, and social equity. The study first evaluates the adequacy of existing regulatory frameworks, including the EU AI Act and Italian legislative reforms, finding significant gaps in addressing algorithmic discrimination in fiscal contexts. Through a theoretical model of taxpayer–tax authority negotiations, the paper demonstrates how AI systems relying on biased training data or latent discriminatory variables can generate market failures in the form of adverse selection and moral hazard. The model reveals that algorithmic discrimination creates perverse incentives: Honest taxpayers from disadvantaged groups may face higher rejection rates for settlement offers due to artificially deflated enforcement probability estimates, while strategic actors exploit these biases to secure favourable terms through manipulative strategies. This leads to Pareto-suboptimal equilibria that reduce fiscal efficiency and exacerbate economic inequality. Findings show that even taxpayers with identical observable characteristics may face unequal treatment when latent bias indicators penalise certain demographic groups, violating principles of vertical equity and administrative impartiality. This research contributes to the literature on AI governance in public finance by integrating economic theory with legal analysis to propose policy solutions that balance technological efficiency with distributional justice. The study concludes with recommendations for developing transparent, auditable, and contestable algorithmic models in tax enforcement, supported by regulatory frameworks codifying fairness principles.