What are the fair housing rules for AI tenant screening?
A score that says decline and cannot say why is the shape HUD warned about. The fix is a product decision, and it is cheaper before the model exists.
The short answer
6 things that decide this
- 01The Fair Housing Act applies to a tenant screening decision whether a person, a vendor or a model makes it, and buying a third-party screening product does not move the liability off the housing provider.
- 02HUD's Office of Fair Housing and Equal Opportunity issued guidance on screening applicants for rental housing on 2 May 2024, and it states that denial recommendations or low grades should not be provided to housing providers in a conclusory fashion.
- 03The guidance says screening reports should include all available relevant details about the basis for the determination, giving dates, locations, case numbers and dispositions as the examples.
- 04Those details should be in plain language, so a lay person can easily understand the reasons for the denial.
- 05Applicants must get a real chance to dispute the accuracy or completeness of negative information, including showing that a record belongs to someone with a similar name or that a court decided in their favour.
- 06Liability does not need intent. A neutral rule with a discriminatory effect is a disparate impact claim, and criminal-record and blanket exclusions are the classic target.
A model discriminates by race without ever seeing race
Teams strip the protected traits out of the training data and believe the problem is handled. It is not. ZIP code, neighbourhood, school district and applicants like our current residents all stand in for race and national origin. In American housing, they always did.
Then there is the label problem, which is worse. Train on your past approvals and the model learns your past decisions, including the ones you would not defend in a deposition. It scores well on every accuracy test, because the historical labels are the bias. Accuracy against a biased record is not a defence. It is the finding.
Catching this needs outcome testing across protected groups. Almost nobody builds it, and it is the one measurement that would have told you.
The guidance, read as a product specification
Each row is a design choice. Cheap on day one, costly once real decisions have been made against the old model.
| What HUD points at | What the product has to do |
|---|---|
| No conclusory denial recommendations or low grades | The screen returns findings with their basis, not a red badge. A grade with no reasons behind it should not be the thing a leasing agent acts on. |
| All available relevant details about the basis | Dates, locations, case numbers and dispositions are stored as fields and shown. That means an arrest can be told apart from a conviction, and a dismissed case from a judgment. |
| Plain language a lay person can understand | Reasons render as sentences, not as internal codes or feature names. Write them for the applicant, who will be reading them upset. |
| Customisability of criteria and weights | The housing provider sets the criteria and the thresholds. The vendor implements the policy rather than setting it, and every criteria set is versioned. |
| The applicant can dispute accuracy or completeness | A dispute intake, an investigation path, a correction, and re-notification of anyone who received the wrong report. This is a first-class flow, not an inbox. |
| Individualised assessment over automatic denial | Mitigating evidence attaches to the application and reaches the decision-maker. A blanket rule cannot be the only path to a decline. |
- CriteriaVersioned, set by the provider.
- InputsSources named, matches scored.
- FindingsDisposition, never a boolean.
- DecisionReasons stored beside it.
- DisputeCorrect, then re-notify.
Store the decision alone and you cannot reconstruct it months later. That failure is fatal twice over: in a HUD complaint, and again in a credit-reporting dispute.
Fair housing is only half of the exposure
Pull third-party data into a report a landlord uses to decide, and you are usually a consumer reporting agency under the Fair Credit Reporting Act. Two sets of rules now govern one screen. They overlap, and they ask for different things.
The requirement first builds miss is the adverse action notice. If a decline rests even partly on a consumer report, the applicant is owed a notice naming the reporting agency and telling them how to dispute it. Build that pipeline before you build the score. Teams do it the other way round and then discover the notice has nothing specific to say, because the model's output was a number.
- 01Log the permissible purpose and the applicant's authorisation with every report pulled.
- 02Filter obsolete information when the report is generated, not when the data arrives.
- 03Require more than a name and an approximate date of birth before a public record attaches to a person.
- 04Never reuse a cached report for a later decision without a fresh purpose.
Related questions
01Is AI tenant screening illegal?
No, and HUD did not ban it. The guidance issued on 2 May 2024 sets expectations for how screening is done, and the Fair Housing Act applies to the outcome regardless of what produced it. An opaque score that outputs a decline with no reasons is the shape that draws attention.
02Does using a third-party screening vendor protect us?
No. HUD's guidance addresses housing providers and screening companies together, so both carry obligations. Ask a vendor which criteria you can set, whether the report shows dispositions, and how an applicant disputes a record. Get the answers before the contract.
03Can we still exclude applicants with criminal records?
Blanket criminal-record bans are the classic disparate impact target, so a flat rule is the risky design. HUD expects an individual assessment rather than an automatic denial. In practice that means the kind of record, how old it is, and anything the applicant offers in their favour all reach a person who can weigh them.
04What should a screening report actually show?
The basis, not the conclusion. HUD's guidance points at dates, locations, case numbers and dispositions, in plain language a lay person can understand. If your UI reduces all of that to a colour, the underlying data is probably fine and the product is the problem.
05How do we know our model is not discriminating?
Measure outcome rates across protected groups and keep measuring after launch. That test is the one almost nobody builds, and it is the only thing that catches a proxy variable doing the work of a protected attribute. Document the search for a less discriminatory alternative while you are at it.
Related
- Real estate software development →What we build in property, and the two systems running today.
- Is algorithmic rent pricing legal? →The other regulated model in a property stack, and what the RealPage decree changed.
- AVM →Automated valuation models, and why a point estimate is the wrong output.
- SmartREI →Investment analysis for US property buyers.
