Is AI Tenant Screening Fair Housing Compliant?

If you are asking "is AI tenant screening fair housing compliant?", you do not need a vendor pitch about unbiased scores. You need a compliance playbook: what the Fair Housing Act actually protects, how disparate impact applies to a model that never sees a photo, what HUD has said about tenant screening and algorithms, when the Fair Credit Reporting Act requires an adverse action notice, and which calls a person still has to make.

This guide is for property-management companies and self-managing owners. It is not a product review and not legal advice. Day-to-day AI jobs live in How Can Property Managers Use AI Day to Day?. The operating loop lives in How Do I Build an AI-First Property Management Workflow?. Copy-paste drafting language lives in Which ChatGPT Prompts Save Hours Every Week?. This post is the screening decision itself: who gets the unit, who gets a denial, and whether a model is allowed anywhere near that call.

Educational only, not legal advice. Fair housing, source-of-income rules, criminal-history limits, and consumer-reporting duties vary by city and state. A tool that "feels consistent" is not a defense. Talk to a local housing attorney before you turn on auto-deny. Never paste Social Security numbers, full credit files, bank statements, or extra identity documents into a general chat tool.

Quick answer: the model can compare a file to my written rules. It cannot decide.

Not automatically. AI tenant screening is fair-housing compliant only if I already have a written, tenancy-related standard, I apply that standard the same way to every complete file, the tool does not use protected-class data or close proxies, a person reviews anything that is not a clean match, I send a real adverse action notice when a consumer report is involved, and I follow the stricter of federal, state, and local rules.

No — if the product ranks people by a hidden score, auto-rejects on an arrest, treats a housing voucher differently from wages, or writes a warmer denial for the applicant who "sounds professional."

Empty days still cost money while I do this carefully. After I lock the written standard, I run the wait through the Vacancy Cost Calculator so I am not tempted to let the software pick because the unit has been sitting.

What I mean by AI tenant screening

I treat four jobs as four different risk levels. Vendors mix them on one invoice.

Completeness check. Did the application include ID, income proof, authorization to pull a report, and every field I require of everyone? Low risk if the list is the same for every applicant.
File-to-criteria compare. Does this file meet the income multiple, rental-history rule, and occupancy limit I published? Medium risk. Useful if the output is a checklist, not a personality score.
Report summary. A short restatement of a credit, eviction, or criminal report I already pulled through a proper consumer-reporting channel. Medium-high risk. The summary can drop context or mix two people with similar names.
Approve / deny recommendation. A model that says "decline" or ranks applicants 1 through 12. High risk. This is the product I do not turn on.

ChatGPT pasted a PDF is not a screening program. A tenant-screening API that returns a color and a sentence is still a decision system if I follow the color. I stay liable either way. HUD has been clear that a housing provider does not step outside the Fair Housing Act by handing the work to software or a screening company.

If I am still building the paper process, I start with How Do I Build a Rental Application? and How Do I Check Tenant References Safely?. The model does not replace those steps. Hommy's Tenant Screening Checklist is the written list I compare against — a checklist, not a score.

Protected classes I cannot screen around

Federal fair housing still applies when a laptop does the sorting. The Fair Housing Act names race, color, national origin, religion, sex, familial status, and disability. HUD interprets sex to include sexual orientation and gender identity. Familial status covers households with children under 18 and people who are pregnant or securing custody. Disability covers physical and mental impairments that substantially limit a major life activity, plus a record of such an impairment and being regarded as having one.

State and city codes often add more. Common extras I check before I write criteria: source of income (wages, Social Security, child support, housing choice vouchers), marital status, age, military or veteran status, immigration status, and criminal-history limits that are stricter than federal guidance. Source-of-income rules are the ones AI products skip most often. If my market treats a lawful voucher as a protected source, a model that only "understands" paystubs is already steering.

I do not let a tool infer any of this from a name, a voice note, a photo, an accent in an inquiry, a school, a church, or a ZIP code. That is still screening. How Do I Screen Tenants Without Legal Risk? is the legal frame for the whole file. This post is only the algorithm sitting on top of it.

Lifestyle language is not a loophole. "Young professional," "perfect for a couple," "quiet building, not for kids," and "we prefer working tenants" are the same problems they are in a listing. Can AI Write Rental Listings That Actually Get Leads? is where I scrub ads. I scrub the screening notes the same way.

Disparate treatment and disparate impact both apply to software

Two theories, one inbox.

Disparate treatment is different rules for different people. A model that asks extra questions after it "hears" a family, or that writes a longer, friendlier follow-up to the applicant with polished English, is treatment. So is a human who overrides the tool only for people they like.

Disparate impact is a facially neutral rule that excludes a protected class without a legally sufficient justification, when a less discriminatory alternative would serve the same substantial, legitimate interest. Credit cutoffs, eviction lookbacks, criminal blanket bans, and "no voucher software" can all land here even if nobody typed a protected class into the prompt.

HUD's tenant-screening guidance treats algorithms as capable of both. A vendor that will not tell me which factors move the score is a vendor I cannot defend. "The AI said no" is not a business necessity. I have to be able to name the criterion, show it is related to tenancy — paying rent, caring for the unit, complying with the lease — and show I considered a tighter rule if a broad one sweeps in people who would have paid the rent.

I do not invent a success rate for my buildings and call that science. I write the smallest rule that actually targets the risk I am trying to avoid, and I apply it on paper first. Whom Should I Reject for My Rental? is the denial playbook. The model does not get a second, secret playbook.

HUD's screening-and-algorithm points I actually operate on

I am not going to quote a memo at a prospect. I am going to run a process that survives one.

What I take from HUD's public guidance on tenant screening, criminal records, and automated tools:

I stay responsible. Buying a report or an "AI score" does not move Fair Housing Act liability off my lease. Screening companies can have their own exposure too. I still own the decision I implement.
Accuracy is part of fairness. Mixed files, common names, sealed records, and old evictions that were dismissed show up in tenant-screening products. If I cannot let the applicant see and dispute what I used, I should not use it to deny.
Arrest is not conviction. A record of an arrest, without more, is a weak predictor and a common disparate-impact problem. HUD has said as much in its criminal-records guidance for years.
Blanket criminal bans are hard to defend. A "no criminal history" rule that ignores nature, recency, and the tenancy-related risk is the kind of policy HUD has flagged. Individualized assessment is the safer habit: what was the conduct, how long ago, what has happened since, and what does it have to do with this unit.
Eviction and credit records are not magic. An eviction filing is not the same as a judgment. A thin credit file is not the same as unpaid rent. Over-weighting either can exclude people who would have performed.
Applicants need a real way to explain. Automated denial with no path to correct a mismatch, show a voucher payment standard, or request a disability-related accommodation is the opposite of an individualized look.
Proxies count. Neighborhood scores, first names, colleges, and "professionalism" ratings are often just protected-class inference with extra steps.

I do not need a statistic to act on this. I need a paper trail that shows I used a tenancy-related rule, I checked the file for the wrong person, and a human read the edge cases.

Written criteria I apply the same way every time

The cheapest fair-housing control is boring: write the standard before I advertise, put it on one page, and refuse to freelance.

My one-pager usually includes:

Occupancy limit tied to legitimate health and safety or local code, not "I don't want kids"
Income or rent-to-income standard that counts lawful source of income the same as wages where the law requires it
Rental-history lookback I can explain, such as unpaid rent or lease-breaking in a set window
Credit rule only if I actually use credit, with a path for thin files and a path that does not punish a voucher household for lacking a FICO story
Criminal-history rule that names categories and lookbacks, not "any record"
Pet policy as a property rule, plus a line that assistance animals follow the law
The documents I require of everyone, and the deadline I give everyone

I do not keep a side list in my head. I do not add a "good vibe" box. Is This Tenant Truly the Right Fit? is about reliable tenancy signals, not a personality contest the model should score.

The Tenant Screening Checklist is where I tick the same boxes on every file: application complete, income, history, reports, decision, and adverse-action notes. If the AI output cannot be pasted next to those boxes, it does not belong in the file.

When income is the gate, I use a number I can show. Hommy's Rent Affordability Calculator is for the math, not for guessing whether someone "seems like they can pay."

What I never feed a model, and what I never let it decide

General chat tools are not consumer-reporting agencies. I do not paste a full credit report, a Social Security number, a driver's license image, or a bank statement into one. If I want a summary, I use the screening company that pulled the report, or I type the few facts I already decided I am allowed to use, with the name stripped.

Never as an input

Photos or video tours of the applicant
Voice notes used to judge "fit"
Prompts like "does this person sound like a good tenant"
School, church, or employer as a quality signal
"Read this inquiry and tell me if I should bother processing the application"

Never as an output I accept

A ranked list of people
Auto-deny
A denial reason the file does not support
A different follow-up question for one applicant
A recommendation to skip the assistance-animal or accommodation process

After-hours inquiry bots have the same failure mode with a smile. How Do I Use AI to Answer Tenant Inquiries 24/7? is the inbox. That assistant points everyone to the same application. It does not pre-sort.

If I am still choosing a small software stack, Which AI Tools Should a Self-Managing Landlord Start With? is the starter kit. I hang screening on the checklist, not on a new "AI tenant score" subscription.

FCRA adverse action is still my job

When I use a consumer report — credit, many tenant-screening packages, some criminal or eviction products sold as consumer reports — and I take adverse action, the Fair Credit Reporting Act has a notice job that a model cannot finish for me.

Adverse action in renting is broader than "I said no." It can include denying the application, requiring a co-signer I would not have required otherwise, raising the deposit, or offering the unit only on worse terms because of the report.

The notice has to do real work: say that I took adverse action, identify the consumer reporting agency by name and contact information, state that the agency did not make the decision and cannot explain it, and tell the applicant they can get a free copy of the report and dispute errors. State law can add more.

An AI paragraph that says "unfortunately you do not meet our criteria" is not that notice. I can ask a model to draft from a template I already approved. I still fill in the agency, the date, the real reason that matches my written standard, and I send it. I keep a copy in the unit folder. How Do I Organize My Rental Records? is the habit. Chat history is not the record.

If the file is a close call, I do not let the tool pick a harsher reason so the letter "sounds stronger." The reason I send is the reason I used.

Criminal history, source of income, and the local overlay

Federal guidance is the floor. The city can be stricter.

Criminal history. I write the categories I will consider, the lookback, and the rule that arrest-only records do not decide the file. I read the actual disposition. I give the applicant a chance to show the record is wrong or that the conduct is old, sealed, or unrelated to tenancy. Some cities and states limit when I may ask, which offenses I may use, and how long I may look back. A national screening API that only knows "hit / no hit" will miss that overlay. I turn those filters off or I do not use the product.

Source of income. Where vouchers or other lawful benefits are protected, I treat the housing authority's payment standard as income I already know how to verify. I do not require extra months of paystubs from a voucher household that I would not require from a W-2 household. I do not let a model flag "government assistance" as risk.

Eviction records. I distinguish a filing from a judgment, and a judgment from a case that was dismissed or settled. I check whether my city limits the use of old filings.

A bad placement is expensive. That does not license a panicked algorithm. The Eviction Cost Calculator is for understanding what a failed tenancy costs — after I have a lawful standard, not instead of one.

Human review, accommodations, and the decision I still sign

I review every file that is not a clean meet-or-miss on the written list. I review every file that mentions a disability, an assistance animal, or a request for a change in how I apply a rule. I review every criminal or eviction hit. I review every thin credit file. I review every time the model says it is "unsure."

Reasonable accommodation is not a screening score. If someone asks for a different way to verify income, a live-in aide, or an exception to a pet rule for an assistance animal, I pause the auto-template and handle it as an accommodation request. I do not ask the model whether the person "seems disabled."

I sign the decision. Literally: my name or my leasing agent's name on the checklist, with the date, the criterion that decided it, and the notice I sent. If I cannot explain the deny in one sentence that matches the one-pager, I do not send it.

When I do approve, I move to the same move-in process for everyone. Hommy's Move-In / Move-Out Inspection Checklist is that layer. Screening is over. The model does not get a second vote at key handover.

A starter week I actually run

I do not buy auto-deny on a Monday.

Day 1. Write the one-page standard. Pull from the lease, local occupancy rules, and the reports I actually buy — not from a vendor default.

Day 2. Load that standard into the Tenant Screening Checklist. Run last month's files through it by hand. If two files would have gotten different answers for the same facts, I fix the page, not the people.

Day 3. Fair-housing and proxy scrub. No lifestyle words. Source-of-income line present. Assistance-animal line present. Criminal rule is not "any record."

Day 4. Adverse action template. Real consumer-reporting-agency fields. A human fills them.

Day 5. If I use AI at all, I use it only to flag missing documents and to compare a file to the written boxes. I send every message. The week after, I still send every denial myself.

If I am mapping who owns leasing versus who owns the legal risk, What Does a Property Manager Do for Me? is the split. Software will not invent it.

Limits of AI on tenant screening

It cannot see the person, and it should not try. A photo, a voice, or a "professional" writing style is not a rental history.

It cannot be the consumer reporting agency. If I did not pull a report through a lawful channel, I do not have a report. A chatbot summary of a PDF I should not have pasted is worse.

It cannot do individualized assessment. Nature, recency, and rehabilitation are judgment. A score hides them.

It cannot keep up with city hall. Fair-chance ordinances and voucher rules change. A national model lags.

It cannot send a valid adverse action notice unless I already wrote the legal parts and I fill the blanks.

It cannot be my system of record. The checklist, the notice, and the unit folder are the record.

It cannot decide who gets the unit. If I catch it adding extra hurdles for one applicant, I turn it off.

Use AI to keep the file complete and the comparison boring. Use a person for denials, accommodations, criminal and eviction hits, and anything the written page does not cover. That split is the whole compliance story.

FAQ

Can I let a screening product auto-deny applicants?

I do not. Auto-deny skips individualized review, accommodations, mixed-file errors, and local overlays. If a vendor ships it on by default, I turn it off or I do not buy the product.

Is a hidden "tenant score" safer than my own credit cutoff?

Usually the opposite. A cutoff I wrote is something I can explain. A hidden score is something I cannot defend as a business necessity, and I cannot tell an applicant what to dispute.

Do I still send an adverse action notice if the AI only "recommended" a deny?

If I used a consumer report and I took adverse action, yes. The software's label does not change the notice duty. I send it, I name the agency, I keep the copy.

Can I use AI to compare two qualified applicants?

I do not rank people. If two files both meet the written standard, I use the lawful tie-break I already published — for example, first complete file — not a model's preference.

What if the report might be about the wrong person?

I do not deny on a maybe. I give the applicant a chance to show the mismatch. Mixed files and common names are exactly why HUD cares about accuracy in screening.

Where do Hommy's tools sit in this workflow?

Next to the decision, not instead of it. Screening checklist on every file, rent affordability when income is the gate, vacancy cost when I am tempted to rush, eviction cost when I am tempted to panic. Open the landlord tools when I need a number or a written list. The model does not get the pen.

Final takeaway

I use AI in tenant screening only as a clerk: complete file, same boxes, draft from a template I already approved. Fair-housing compliance is not a feature a vendor can toggle. It is a written standard, the Fair Housing Act's protected classes, an honest look at disparate impact, HUD's screening-and-algorithm cautions, a real FCRA notice, local source-of-income and criminal-history rules, and a human who still signs the deny.

The model can save me from a missing paystub. It cannot save me from a cheerful, automated "you don't qualify" that I cannot explain.

When I am ready to run the same list on the next file, I open the Tenant Screening Checklist and I keep the decision boring on purpose.