Fairness in Automated Tenant Screening Systems
Keywords:
automated tenant screening, algorithmic fairness, housing discrimination, responsible AI, responsible, fair housing, explainability, human-in-the-loop, adverse decisionsAbstract
Automated tenant-screening processes increasingly characterize housing applicants based oncorings derived from combinations of credit, rental history, court records, criminal-history content, income indicators, and specialized risk scores. Automation can preserve speed and replicability but may perpetuate past discrimination, reinforce errors that appear in record matching, interdict explanation and applicant feedback, and undermine due-process. Job-related predictive information alone cannot define fair tenant screening; reachalonepredictive effectiveness and overlookbankrupt approval-rate parity provide an incomplete picture. This study devises a simulation framework forapiecepredicative-utility, false-denial-gap, residual-group-error, explainability, data-minimization, human-in-the-loop, auditability, andappeal.
A synthetic data-generating process was used to deliver 2,400 fictive rental applications equally among four simulated architectures: AI-Only, Explainable AI, Fairness-Aware AI, and Responsible AI. No real-life childless-and-oversexed applicant, landlord, borrow-from-cash, bring-to-eviction, criminal-history, disability, race, income, or consumer-report data entered the model. The Responsible-AI configuration maintained an overall predictive accuracy of 89.7% while homogenizing the modeled false-denial gap between synthetic comparison groups from 8.8 percentage points in the baseline-no fairness architecture to 0.9 percentage points in the final structure. Human read-aide decreased the reproduced number of over-furloughed high-risk false positives most in cases featuring roundabout eviction and criminal-history content. Linear-regression analysis identified unresolved record ambiguity and proxy-rich scoring as major positive-housing-risk predictors, while audit-friendly
applicant explanation, data-minimization, individualized review, and appeal were negative predictors.
The author concludes that fair hire selection necessitates a sociotechnical governance framework instead of a single mathematical fairness constraint. Automated scoring must be nested in a process that checks data quality, allows explanation, enables meaningful appeal, analyzes system-wide subgroup errors, and prevents nearly-uncertain scores from eventually becoming unchallengeable housing denials.
