# Fair housing and discrimination risk in algorithmic decisions
In 2019, a tenant screening company called SafeRent Solutions generated a score for a Massachusetts woman applying to rent an apartment with a housing voucher. Her score was low. She was denied. She had no eviction record and a strong voucher-backed payment history covering her rent. The algorithm did not know her race. It did not need to. A 2024 settlement in *Louis v. SafeRent Solutions* alleged the scoring model relied on factors, including credit history and non-relevant transaction data, that produced disproportionately lower scores for Black and Hispanic applicants and voucher holders as a group, even though the protected characteristics themselves were never inputs. That is disparate impact, and it is the central compliance trap in real estate AI.
This lesson unpacks how it happens, which laws apply, and what checks a real estate organization must run before letting a model touch a leasing, underwriting, or pricing decision.
The Fair Housing Act (FHA), enacted in 1968 and enforced by the U.S. Department of Housing and Urban Development (HUD) and the Department of Justice (DOJ), prohibits discrimination in housing on the basis of race, color, national origin, religion, sex, familial status, and disability.
Two liability theories matter for AI:
State and local overlays add complexity: California, New York, and Colorado all have their own fair housing and AI-specific disclosure statutes, and some cities (Colorado state law effective 2026, for instance) require impact assessments before deploying "high-risk" automated decision systems in housing.
A proxy variable is a neutral-looking input that correlates strongly with a protected class. Real estate models are full of them:
None of these variables mention race. All of them can reconstruct race statistically. A model trained to minimize default risk will happily pick up zip code or eviction flags because they are predictive, without any awareness that they are also discriminatory proxies.
Model risk here is not "the model is inaccurate." The model can be highly accurate at predicting late payment and still be illegal, if its accuracy is achieved by leaning on variables that track protected class membership more than they track actual tenant risk. Accuracy and fairness are separate axes. A governance program that only checks predictive performance (AUC, precision, recall) will miss this entirely.
1. Training data bias: historical acceptance/rejection data already embeds past discriminatory decisions (human or algorithmic). Training on it reproduces the bias.
2. Proxy leakage: as above, neutral features carry protected class signal.
3. Vendor opacity: most landlords do not build screening models in-house. They buy from vendors (SafeRent, RealPage, CoreLogic, TransUnion SmartMove are commonly used names in this space). If the vendor will not disclose model logic or validation data, the landlord still bears legal liability as the entity making the housing decision.
4. Score-threshold rigidity: automated cutoffs ("score below 500, auto-reject") remove human discretion that might otherwise catch an unfair edge case, and they scale a single flawed rule across thousands of applicants at once.
5. Feedback loops: a model that suppresses applications from certain zip codes reduces future data from those areas, reinforcing its own blind spots over time.
A pre-deployment fair housing check for any tenant-facing or pricing-facing AI system should include:
1. Disparate impact testing. Run the model's outcomes against applicant demographic data (collected separately, never as a model input) using the "four-fifths rule": if the selection rate for a protected group is less than 80% of the rate for the highest-scoring group, that is a common (though not conclusive) regulatory red flag used by the EEOC and referenced in HUD guidance.
Simple worked example:
2. Proxy correlation audits. Statistically test each input feature against protected class membership (using indirect data such as Bayesian surname/geography imputation, a common technique regulators themselves use, called BISG, Bayesian Improved Surname Geocoding). High correlation does not automatically mean removal, but it triggers a business-necessity review.
# Simplified proxy correlation check
import pandas as pd
# df has model_score, zip_code, and a separately-held
# race_proxy column (never fed into the model)
corr = df.groupby('zip_code')['race_proxy_pct_minority'].mean().corr(
df.groupby('zip_code')['model_score'].mean()
)
print(f"Zip-code to score correlation: {corr:.2f}")
# A strong correlation (e.g., > 0.5) warrants a business-necessity
# justification and a less discriminatory alternative test3. Less discriminatory alternative (LDA) testing. Fair lending regulators (and by extension, fair housing practice) expect firms to test whether a model with similar accuracy but lower disparate impact exists. If your data science team can build an alternative with a smaller 80% rule gap and comparable predictive power, deploying the more discriminatory version is a legal exposure, even if the first model is technically "better."
4. Vendor due diligence. Before signing with any screening or pricing vendor, request: their most recent disparate impact test results, the feature list, and whether they will indemnify you. HUD and the Consumer Financial Protection Bureau (CFPB) have both signaled that landlords cannot outsource liability by pointing at the vendor.
5. Human review and appeal path. No fully automated denial without a human-reviewable override. This is best practice and, in some states, becoming a legal requirement for "high-risk" automated decisions.
Knowledge check
1. In the SafeRent tenant screening case, the scoring model never used race as an input, yet the outcome allegedly harmed Black and Hispanic applicants disproportionately. Which liability theory does this illustrate?
2. Why is disparate treatment liability rare in modern AI-driven housing cases, while disparate impact is the primary compliance risk?
3. A real estate company is deciding whether to deploy a new applicant scoring model. Based on the *Inclusive Communities* precedent, what should this most directly prompt the company to do?
4. Select ALL correct answers about how proxy variables can create disparate impact risk in tenant screening algorithms.
Select all the correct answers.
5. Select ALL correct answers about the legal landscape governing AI-driven housing decisions.
Select all the correct answers.
Tenant screening is not the only exposure. Algorithmic rent-setting tools (RealPage's YieldStar is the most litigated example, subject of a DOJ antitrust suit filed in 2024 alleging coordinated pricing among landlords) raise a related but distinct issue: antitrust risk from shared pricing data, plus a fair housing angle if the pricing model's inputs correlate with tenant demographics in ways that produce disparate cost burdens. Governance teams should treat pricing and screening as two separate model risk categories, each needing its own testing protocol.
How Algorithms Can Discriminate (and What to Do About It)
For a deeper primer on disparate impact methodology, HUD's own 2013 discriminatory effects rule is the primary source document: HUD Discriminatory Effects Standard.