Why This Use Case Carries Unusual Risk
Tenant screening AI sits at a uniquely difficult intersection. The use case has clear business value: operators want to predict which applicants will be reliable tenants, and AI can produce more accurate predictions than traditional rule-based screening. The legal exposure is also significant. Tenant screening decisions affect housing access, which is regulated under federal and state fair housing law. The intersection has produced multiple multi-million-dollar settlements through 2023, 2024, and 2025.
The settlements have established specific patterns of what regulators and courts find problematic. Models that rely on criminal history in ways that produce disparate impact. Models that use proxies for protected class without justification. Models that have not been tested for disparate impact. Models that lack documentation of business necessity for features that produce disparities.
A counsel I worked with at a tenant screening provider described the lesson their industry has learned. "We used to think model performance was the metric. We learned that disparate impact is the metric, and that performance is constrained by what we can defend legally. The constraint changed the engineering." The framing has stayed with me.
The patterns for tenant screening AI that work in 2026 reflect this learning. They are not the patterns of typical machine learning model development. They are model development plus a layer of legal constraint that shapes what is acceptable.
EHR Integration Reality
The three gaps between Epic's FHIR R4 documentation and production behavior.
What the Settlements Have Established
Several specific patterns have emerged from the tenant screening settlements through 2023-2025. Operators and vendors deploying tenant screening AI should understand these.
Blanket criminal history exclusions produce disparate impact. The HUD guidance has been clear for years; the enforcement has been increasingly active. Screening models that automatically exclude applicants based on criminal history without individualized assessment face significant legal exposure. The CoreLogic case in 2024 settled for substantial damages on this issue.
Eviction history use requires care. Eviction filings (as opposed to judgments) include cases that were dismissed or resolved in the tenant's favor. Models that treat all eviction filings as negative produce disparate impact. The pattern requires distinguishing filings from judgments and considering the context.
Income source restrictions face fair housing scrutiny. Section 8 voucher discrimination is prohibited in many jurisdictions. Models that effectively discriminate by income source through other proxies face legal exposure even where source-of-income discrimination is not explicitly prohibited at the federal level.
Geographic proxies produce disparate impact. ZIP codes, neighborhoods, and other location-based features can act as proxies for race and ethnicity. Models using these features without disparate impact testing produce legal risk.
The cumulative impact of these patterns is that tenant screening models in 2026 cannot be developed the way models for many other use cases can. The legal constraints are part of the model architecture, not an afterthought.
What Defensible Tenant Screening AI Looks Like
The patterns that work in 2026 share specific properties. The properties reflect the legal environment rather than just technical considerations.
Individualized assessment for high-stakes signals. Criminal history, eviction history, and similar signals get individualized consideration rather than blanket exclusion. The patterns that work include consideration of time since the event, severity, relevance to tenancy, and rehabilitation evidence.
Business necessity documentation for every model feature. Each feature that influences the screening decision has a documented business necessity tied to tenancy obligations. Features that cannot be tied to legitimate business necessity get removed regardless of predictive value.
Disparate impact testing throughout the model lifecycle. Pre-deployment testing against protected class estimates. Continuous monitoring of production decisions. Adjustments when monitoring reveals drift. The testing is part of the operational practice rather than a one-time exercise.
Transparency to applicants about screening decisions. Adverse action notices that comply with the Fair Credit Reporting Act and similar state laws. Specific reasons rather than generic explanations. The transparency supports both compliance and applicant trust.
Human review for borderline cases. Models produce scores; humans make final decisions on borderline cases. The pattern provides judgment that pure model decisions cannot provide.
Counsel review of model design, testing, and deployment. The legal review is integrated rather than reactive. The patterns that have produced settlements have generally lacked this integration.
The Feature Engineering Challenge
Feature engineering for tenant screening AI requires understanding both what features predict tenancy outcomes and what features create legal exposure. The two considerations sometimes pull in different directions.
Credit-related features have the best track record. Credit scores, payment history, and similar features have been used in tenant screening for decades. The features have predictive value and have been tested in fair housing contexts. The patterns that work use them carefully but extensively.
Income-related features are central but require care. Income-to-rent ratios. Income stability indicators. Employment history. The features are predictive. The use has to avoid effective source-of-income discrimination and has to consider whether the income thresholds produce disparate impact.
Behavioral features from prior tenancies have value but are difficult to source consistently. Payment timing patterns. Lease compliance. Maintenance interaction patterns. The data exists in property management systems but is fragmented across operators and platforms.
Public record features require careful handling. Court records. Bankruptcy filings. Tax liens. The features can be predictive but raise both accuracy issues (records can be inaccurate or out of date) and fair housing issues (the records can carry disparate impact). The handling matters.
Demographic and demographic-adjacent features should not be used directly. Race, ethnicity, religion, family status, disability, national origin. Direct use is disparate treatment. Proxies for these create disparate impact. The features and likely proxies should be reviewed and removed.
Monitoring and Ongoing Operation
Tenant screening AI requires monitoring practices that produce defensible evidence of ongoing compliance. The monitoring is part of the model rather than a separate activity.
Disparate impact monitoring runs continuously on production decisions. The monitoring uses BISG or similar methods to estimate protected class for applicants. The aggregate decision outcomes are tested for disparate impact. The monitoring catches drift that pre-deployment testing missed.
Outcome monitoring tracks the screening's predictive accuracy. The model's predictions should match observed outcomes. Drift in accuracy may indicate that the model is no longer working as intended. The monitoring supports continuous improvement.
Adverse action monitoring tracks how applicants respond to adverse decisions. Disputes. Re-applications. Litigation. The patterns surface issues with the model or the explanations that the broader monitoring may miss.
Periodic external review provides independent assessment of the model and its operation. The review may be conducted by counsel, by specialized compliance consultants, or by third-party auditors. The frequency depends on the use case risk and regulatory requirements.
Incident response procedures handle the cases when monitoring reveals problems. Disparate impact above thresholds. Predictive accuracy degradation. Applicant complaints suggesting systemic issues. The procedures are documented and practiced.
What Modern Tenant Screening AI Looks Like
The reference patterns in 2026 share recognizable components across tenant screening providers and operators that have built defensible practices.
Individualized assessment for criminal history, eviction history, and similar high-impact signals. Blanket exclusion patterns have been phased out.
Business necessity documentation for every feature. The documentation supports the legal defense of the model design.
Continuous disparate impact monitoring with documented methodology. The monitoring catches drift and supports regulatory engagement.
Transparency to applicants through clear adverse action notices and accessible explanations. The transparency builds applicant trust and supports compliance.
Human review for borderline cases and for cases that fall outside the model's confidence range. The judgment is part of the system.
Counsel integration throughout. The legal review is continuous rather than reactive. The model decisions are legally defensible by design.
The patterns are not specific to any single vendor. They apply across the tenant screening industry. Operators using vendors should evaluate the vendors against these patterns. Vendors providing tenant screening services should be developing toward these patterns.
HIPAA and AI
Why 90% of healthcare organizations are unknowingly exposing patient data through AI tools.
What Logiciel Does Here
Logiciel works with tenant screening providers and multifamily operators building defensible AI screening systems. The work is typically structured around model design, disparate impact testing infrastructure, and counsel-integrated governance alongside the AI development.
The AI Governance for Fair Housing framework covers the broader patterns. The AI Reliability framework covers the monitoring and testing infrastructure that tenant screening AI requires.
A 30-minute working session is enough to assess your current tenant screening AI against the defensible patterns.
Frequently Asked Questions
Can we use AI for tenant screening at all given the legal risk?
Yes, with the right practices. The risk comes from how the AI is built and operated rather than from using AI per se. The vendors and operators with defensible practices continue to use AI successfully. The ones with poor practices have produced the settlements that get attention.
Should we use protected class data directly in disparate impact testing?
Most operators do not collect protected class data directly because of legal and ethical concerns. BISG (Bayesian Improved Surname and Geocoding) and similar estimation methods are used for aggregate disparate impact analysis. The CFPB has used BISG in regulatory contexts; it is an accepted methodology.
What is the role of human review in AI screening?
Significant. Models produce scores; humans should make final decisions on borderline cases, on cases involving high-impact signals (criminal history, eviction history), and on cases where the model's confidence is low. Pure automated decisions reduce the defensibility and create unnecessary legal exposure.
How do we handle applicants with criminal histories?
Individualized assessment. The HUD guidance has been clear: blanket exclusions produce disparate impact. The assessment considers the nature of the offense, time since the offense, the applicant's rehabilitation, and the relevance to tenancy obligations. The assessment can be supported by AI; it should not be replaced by AI.
What about state-specific tenant screening laws?
Significant variation. California, New York, Washington, and several other states have specific tenant screening requirements that go beyond federal fair housing law. Source-of-income protection. Criminal history limitations. Specific disclosure requirements. Operators in multiple states have to handle the variation, often through state-specific model configurations. ## Sources: HUD Fair Housing Act Enforcement Guidance, 2024 Connecticut Fair Housing Center v. CoreLogic, 2024 CFPB Algorithmic Lending Guidance, 2023
