SSRN Review & Roundup: Eyal Reviews Mazur’s Modernizing the IRS in the Age of AI

This week, Mirit Eyal (Alabama; Google Scholar) reviews Orly Mazur (SMU, Google Scholar), Modernizing the IRS in the Age of AI, 45 Va. Tax Rev. 182 (2025):

Artificial intelligence (“AI”) is no longer an experimental technology confined to labs and tech firms. It is shaping government operations in real time. Federal agencies are deploying generative models to draft correspondence, using anomaly detection to flag fraud, and testing predictive systems to allocate limited enforcement resources. The pressure to modernize is growing, and the political narrative often frames AI as both inevitable and transformative. Against this environment, this article functions as a reality check. Mazur asks whether the IRS, as currently constituted, possesses the institutional capacity, technological infrastructure, and governance frameworks necessary to deploy AI responsibly. Her answer is a cautious, but evidence-based no.

Mazur’s core insight is that institutional capacity, not technological sophistication, determines whether AI strengthens or weakens public administration.

The IRS possesses an extraordinary mandate. It must collect trillions in revenue each year, manage over 260 million returns, adjudicate complex disputes, detect sophisticated evasion schemes, and serve as the primary point of contact between citizens and the federal state. AI could, in theory, enhance each of these functions. Mazur acknowledges this and carefully describes the areas where AI could indeed be transformative. For example, generative systems can summarize large document sets for auditors, extract relevant information from mixed structured and unstructured data, and draft routine notices with greater consistency. Machine learning can detect filing patterns that historically require specialized human review.

However, Mazur does not treat these possibilities as simple solutions. She grounds her analysis in the technological landscape the IRS currently inhabits. The agency still depends on the Individual Master File, a system designed in the 1960s that lacks interoperability with modern data pipelines. It operates a patchwork of systems that store data in incompatible formats. Some records are digitized through optical character recognition, others are manually keyed in, and still others are scattered across legacy databases. AI cannot overcome these inconsistencies because AI depends on consistent and reliable data. A model trained on fragmented, poorly labeled, or outdated information will replicate those flaws with greater speed and scale. The point is not that AI performs poorly. It is that the IRS lacks the foundation necessary for AI to perform well.

Mazur advances a related claim about the limitations of existing automated systems the IRS already uses. These systems reveal what happens when automation takes root without structural reform. The Automated Underreporter Program illustrates how rule-based systems generate huge numbers of mismatches that do not result in actual noncompliance. The large partnership audit models were built using incomplete data that skewed risk assessments. These failures are not hypothetical. They are documented outcomes that expose the risks of adopting more complex AI systems without addressing data quality and governance. If relatively simple algorithms generate high false positive rates, there is no reason to assume that more complex systems will be more reliable.

Her discussion of bias is equally substantive. Mazur connects the disproportionate audit rates experienced by low-income and Black taxpayers to the structural incentives embedded in current enforcement patterns. The IRS devotes significant resources to areas where it believes detection is more likely and where cases are cheaper to resolve. These structural incentives lead to audit patterns that fall heavily on EITC claimants even though the majority of the tax gap resides in high-income business owners and complex pass-through structures. AI systems trained on historical audit data will learn these patterns. They will reproduce the agency’s structural biases because they reflect the underlying data environment, not model error. Mazur presses the point that bias is not merely a characteristic of algorithms. It is embedded in institutional choices that have shaped the data over time. An agency that has never conducted a systematic demographic analysis of its datasets is not prepared to deploy AI that amplifies the patterns within them.

Mazur also deepens the conversation about transparency by emphasizing that opacity is not solely a technical limitation but a governance challenge. When an automated system flags a return, the IRS must be able to explain why. Taxpayers have statutory rights that require the government to provide reasons for its determinations. Mazur argues that black box systems complicate this obligation. Moreover, if the agency itself cannot fully understand how a model generated its output, accountability erodes. This raises fundamental administrative law concerns that go beyond the IRS. Any agency deploying AI must retain the capacity to justify and review its actions. Mazur warns that the allure of automation can tempt agencies to rely on tools that outpace their capacity for oversight.

Her historical analysis further reinforces the structural critique. The IRS has spent decades attempting to replace obsolete systems, yet these efforts have repeatedly collapsed under unstable funding cycles, political interference, and shifting priorities. The result is a modernization landscape that is incomplete, disjointed, and fragile. Mazur situates AI within this history to argue that the agency is trying to add advanced capabilities without resolving foundational deficiencies. The lesson is not pessimistic. It is pragmatic. Modernization requires reliability, continuity, and long term commitment. The IRS has rarely been given these conditions, and AI cannot create them.

Mazur’s roadmap for reform is therefore substantive rather than aspirational. She calls for stable multi-year funding because modernization cannot proceed in two-year increments tied to political cycles. She argues for workforce renewal because AI requires staff who can evaluate, supervise, and audit model behavior. She emphasizes the need for structured decommissioning of legacy systems because new tools should not be forced to depend on old architecture. She also stresses the necessity of enterprise-wide governance to coordinate AI adoption across divisions that historically operate in silos. These recommendations are not theoretical improvements. They are the basic prerequisites for deploying any high-risk technology in a system as large and complex as the federal tax administration.

There are also deeper policy implications in Mazur’s argument. Her analysis suggests that AI adoption is not primarily a technological project but a legal and institutional one. It requires decisions about how to preserve due process, how to incorporate explainability into enforcement, how to document model behavior for audit purposes, and how to align algorithmic outputs with statutory obligations. Other jurisdictions have begun to develop formal algorithmic accountability regimes. The IRS has not yet done so. Mazur’s article implicitly invites lawmakers to consider how such frameworks should apply in the tax context. Her work also opens the door to comparative analysis. Jurisdictions like Australia and the United Kingdom have had more sustained investment in tax technology and may offer models for phased modernization. Countries that have adopted real time withholding and integrated data systems have a technological baseline that makes AI more viable. Mazur’s argument suggests that the United States cannot skip these foundational steps.

There are also several meaningful opportunities for the article to push its analysis further. One concerns Mazur’s discussion of algorithmic opacity, which identifies the problem clearly but could be strengthened by engaging more directly with concrete transparency mechanisms that are now emerging across administrative agencies. For example, the article could explore how tools such as mandatory model documentation, internal “model cards,” and standardized algorithmic impact assessments might be adapted to the tax context. These mechanisms would not merely promote abstract transparency, but would create enforceable records that allow the IRS, courts, and taxpayers to understand what data inputs were used, how risk factors were weighted, and where human discretion intervened. Similarly, the article could examine whether explainability standards should be tiered, with heightened justification requirements when automated tools influence audit selection or penalty determinations that directly affect taxpayer rights.

A second area for deeper development involves the operational realities of modernization. While the roadmap rightly emphasizes sequencing and structural reform, it could further grapple with the practical challenge of transforming core systems while the IRS must continue processing hundreds of millions of returns on rigid annual cycles. Modernization does not occur in a vacuum. Filing seasons cannot pause, and transitional failures can have immediate downstream consequences for refunds, compliance, and public trust. A more explicit discussion of phased implementation, parallel system operation, and risk management during transition periods would sharpen the roadmap and better reflect the institutional constraints under which the Service operates.

Finally, the article could benefit from stronger engagement with administrative law principles that govern agency decisionmaking. As automated tools increasingly shape audit selection, compliance screening, and enforcement priorities, questions of procedural fairness, reason-giving, and reviewability become unavoidable. The article could more directly address how hybrid human AI systems can be structured to preserve notice, opportunity to respond, and meaningful explanation, particularly where algorithmic outputs influence coercive state action. Clarifying how traditional doctrines of due process and administrative accountability apply in an era of automated enforcement would not only strengthen the article’s legal grounding but also reinforce its broader claim that AI adoption is as much an institutional and legal challenge as it is a technological one.

Nonetheless, Mazur’s contribution is definitive at a moment when AI is shaping the trajectory of federal administration. She provides a nuanced and evidence-rich account of what responsible AI deployment would require in the tax system. Her contribution lies in reframing the question from whether AI could help the IRS to whether the IRS is prepared for the responsibilities that come with AI. She shows that there is no shortcut to the foundational work of modernization, data governance, and accountability. If these conditions are met, AI could yield enormous benefits. If they are not, the technology will amplify the weaknesses of the institution that adopts it.

Here is the rest of this week’s SSRN Tax Roundup:

Syed Shaharyar Ahmed (Ibrar Law), Cross-Border Surveillance and The Right to Privacy: Legal Remedies in the Age of 5g and Iot (Jan. 2026)

Ayomide Alabi (Trellis Limited), Nigeria’s New Tax Regime (2025): Institutional Consolidation, Administrative Ambition, and the Limits of Legal Design (Jan. 2026)

Christina Allen (Curtin U.) & Dale Boccabella (New South Wales), “In Relation to”—The Connection Between Expenditure and Business Under Section 40-880 (Jan. 2026)

Christina Allen (Curtin U.), E-Invoicing for GST Reporting in Australia: Towards Mandatory Implementation (Jan. 2026)

David Okiki Amayo Jr. (U. of Westminster), 3.1: A Jurisprudential Treatise on Continental Sovereignty and Algorithmic Governance (Jan. 2026)

Harald Amberger (Vienna), Ruby Doeleman (Vienna) & Stefanie Pendl (Vienna), (Mis)measurement of Income Shifting (Jan. 2026)

Carlos Barreto (Brawijaya), Unti Ludigdo (Brawijaya), Wuryan Andayani (Brawijaya), & Mohamad Khoiru Rusdy (Brawijaya), Determining Factors of Timor-Leste Tax Revenue Optimization (Jan. 2026)

Andrew Blair-Stanek (Maryland), Nils Holzenberger (Institut Polytechnique de Paris) & Benjamin Van Durme (Johns Hopkins), ChatGPT Generates a Novel Tax Strategy (Jan. 2025)

Eloise Brouillard (U. of Sherbrooke), Policy Forum: The Basis of Customs Duties and Implications of Tariff Tensions, 73 Can. Tax J. 333 (2025)

Anh Nguyen Cao (Nguyen Tat Thanh U.), Interest Tax Shields and Financial Structure: An Accounting-Institutional Diagnosis of State Consistency (Jan. 2026)

Mark J. Cowan (Boise State), Scenes from a Profession (Jan. 2026)