This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A study by researchers from Stanford, the University of Michigan, the University of Chicago, and the Treasury Department found that IRS audit selection algorithms disproportionately target Black taxpayers. Despite being race-blind, the algorithms prioritize low-cost, simple audits, which inadvertently leads to higher audit rates for Black taxpayers compared to non-Black taxpayers. The researchers demonstrated that alternative algorithmic designs could mitigate this disparity without sacrificing tax revenue.
The IRS was auditing Black taxpayers more frequently than other groups allegedly due to the design of their algorithms, focusing on easier-to-conduct audits which inadvertently correlated with the group's pattern of tax filing errors.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Unequal treatment of individuals or groups by AI, often based on race, gender, or other sensitive characteristics, resulting in unfair outcomes and unfair representation of those groups.
Also classified as
The algorithm's effectiveness in identifying actual tax underreporting magnitude differs by demographic group, performing worse at finding high-value evasion when targeting lower-income Black taxpayers.
The IRS audit selection algorithm results in unfair discrimination and unequal outcomes, disproportionately targeting Black taxpayers for audits at rates 2.9 to 4.7 times higher than non-Black taxpayers.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The racial disparity was an unintended consequence of the IRS's deployed audit selection algorithm, which was designed to optimize cost-efficiency and detect underreporting.
Added categories peak at 2: Minor (Psychological Harm, Inferred)
An IRS audit selection algorithm was found to disproportionately target Black taxpayers due to technocratic design choices that optimized for low-cost audits. While raising significant civil rights and administrative equity concerns, the incident has minor national security implications and is manageable through standard domestic policy and algorithmic reforms.
The reports describe the IRS using 'automated systems to select returns for audit' and 'computer algorithms the agency uses to determine who is selected'. These systems run independently to flag returns and generate audit letters (correspondence audits) without individual human agent intervention.
High Risk: The system is used by a government agency for resource allocation and enforcement (tax audits), which directly impacts the fundamental rights and financial well-being of citizens, fitting the category of high-risk AI systems used in law enforcement or essential public services.
The primary goal of the IRS algorithm is to allocate enforcement resources (audits) efficiently to detect tax underreporting and maximize compliance.
The reports are based on a rigorous, peer-reviewed scientific study conducted by major academic institutions (Stanford, Michigan, Chicago) and the Treasury Department, analyzing a massive dataset of 148 million tax returns. The findings are highly consistent across all sources.
None. The statistical disparity is robustly demonstrated, and the IRS itself acknowledged the need to address the findings.
While the exact proprietary code of the IRS algorithm is secret, the researchers successfully modeled its behavior with high confidence, meaning no critical information is missing to classify the safety risk.
Technology
Algorithmic Selection Bias
Algorithm prioritizes audit likelihood over the magnitude of underreporting.
Targeting Refundable Credits
Algorithm targets refundable credit errors rather than total tax evasion.
Dependent Database Flags
Database automatically flags potential issues and generates mail audits.
Data Inputs
Lack of Race Data
IRS does not collect race data, preventing detection of systemic bias.
Race-Blind Model Assumption
Assuming race-blind inputs prevent racially biased algorithmic outcomes.
Simplified Return Focus
Algorithm focuses on simple returns with easy-to-identify errors.
Human Factors
Assumption of Neutrality
Staff assumed race-blind algorithms would produce equitable outcomes.
Overreliance on Automation
Officials relied on automated systems due to resource constraints.
Process and Methods
Cost-Saving Audit Selection
IRS avoids auditing business income returns because they cost more to audit.
Focus on Correspondence Audits
Mail audits are prioritized over complex, resource-intensive field audits.
Minimizing No-Change Rate
Process designed to minimize audits that result in no tax change.
Management
Secret Selection Algorithms
Keep selection algorithms confidential, preventing external equity reviews.
Resource Allocation Choices
Management prioritized cheap, simple audits over high-value complex ones.
Lack of Equity Assessment
Management did not proactively analyze audit algorithms for disparate impacts.
Regulatory Environment
Budget Cuts and Constraints
Congressional budget cuts forced IRS to rely on cheap automated audits.
Lack of Algorithmic Oversight
No regulatory requirements existed to audit federal algorithms for racial bias.
Congressional Focus on Credits
Congressional pressure to prevent over-claiming of refundable tax credits.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
