US housing algorithms: what HUD said on AI and what still applies

In this article

Written by

Kooky

Builder of Shaka, the payment router that pays every agent their commission on closing date.

About Kooky and Shaka →

Software now sits between a rental applicant and a set of keys, and between a housing advert and the people who see it. A screening product may score an applicant before any person reads the file. An advertising system may decide which households are shown a listing and which never learn it exists. In the United States, the federal Fair Housing Act was written long before either tool, yet it governs both.

On 29 April 2024 the Department of Housing and Urban Development (HUD) issued two guidance documents explaining how it read the Act for these tools. Since then the picture has moved: one of the two documents appears in a formal withdrawal notice, the regulation both relied on is the subject of a proposed repeal, and the statute itself is unchanged. This guide sets out what the two documents said, what the law and the federal regulations require with or without them, and exactly what the official pages show about their status as of 10 October 2026.

HUD guidance of 29 April 2024; Federal Register notice of 6 April 2026; Federal Trade Commission guidance for landlords.

Why HUD wrote about algorithms at all

HUD announced the two documents in press release No. 24-098, dated 2 May 2024 and headed "HUD Issues Fair Housing Act Guidance on Applications of Artificial Intelligence". The release says the documents responded to the presidential Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence of 30 October 2023, which asked HUD to address discrimination enabled by automated or algorithmic tools used in housing and real estate decisions.

Both documents came from HUD's Office of Fair Housing and Equal Opportunity, known as FHEO, and neither names an individual signer. The first is the Guidance on Application of the Fair Housing Act to the Screening of Applicants for Rental Housing, which runs to 24 numbered pages. The second is the Guidance on Application of the Fair Housing Act to the Advertising of Housing, Credit, and Other Real Estate-Related Transactions through Digital Platforms.

Neither document created a new rule. Each described how existing law applied to a technology and then listed practices HUD recommended. That distinction matters for everything that follows, because a recommendation can be withdrawn by the agency that made it, while a statute cannot.

What the Fair Housing Act requires, whatever the technology

The Fair Housing Act is found at title 42 of the United States Code, sections 3601 to 3619. As the advertising guidance summarises it, the Act prohibits discrimination in the sale, rental and financing of housing and in related services on the basis of race, colour, religion, sex, national origin, familial status or disability. HUD's current overview page lists the same seven characteristics.

The screening guidance put the central point in one sentence: "The Fair Housing Act applies to housing decisions regardless of what technology is used." A decision to refuse an applicant is a housing decision whether a leasing agent makes it or a model does.

Two legal routes to liability run through both documents. The first is intentional discrimination. In the screening guidance this includes using a protected characteristic, or a proxy for one, as a criterion; adopting overbroad screening aimed at keeping a group out; applying criteria inconsistently; and giving unequal help to applicants. The advertising guidance adds that intentional discrimination covers the use of protected characteristics or their proxies even when an automated system makes the decision.

The second route is discriminatory effect, also called disparate impact, where a practice that is neutral on its face falls more heavily on a protected group. Here the two documents leaned on a federal regulation, section 100.500 of title 24 of the Code of Federal Regulations, and on the Supreme Court's 2015 decision in Texas Department of Housing and Community Affairs v. Inclusive Communities Project, which held that disparate impact claims can be brought under the Act.

For advertising there is also a specific statutory ban. Section 3604(c) of title 42 bars housing notices, statements and advertisements that indicate a preference, limitation or discrimination based on a protected characteristic. The advertising guidance says liability under that provision can be judged by how an "ordinary reader" would understand the advert, regardless of the advertiser's intent.

The three-step test in the federal regulation

Section 100.500 is headed "Discriminatory effect prohibited". According to the text shown by the Electronic Code of Federal Regulations, a practice has a discriminatory effect where it actually or predictably results in a disparate impact on a group of persons, or creates, increases, reinforces or perpetuates segregated housing patterns, because of a protected characteristic. The same section says such a practice may still be lawful if it has what the regulation calls a legally sufficient justification.

That justification has two parts. The practice must be necessary to achieve one or more substantial, legitimate, nondiscriminatory interests of the party defending it, and those interests must not be capable of being served by another practice with a less discriminatory effect. The justification must be supported by evidence; it may not be hypothetical or speculative.

The regulation then divides the burden of proof into three stages.

The regulation closes with a limit: showing a legally sufficient justification is no defence to a claim of intentional discrimination. The test belongs to effect cases only.

What the tenant screening guidance said

The screening guidance was addressed mainly to housing providers and to tenant screening companies, with applicants as a third audience who could use it to understand their rights. It described screening companies as businesses that gather and analyse information about applicants from many sources and produce reports, frequently with a recommendation on whether to accept. Housing providers use those reports to accept, deny or accept with conditions, and some use them for lease renewals. Citing a 2022 Consumer Financial Protection Bureau market report, the guidance put the industry at hundreds of companies earning about US$1 billion a year.

On machine learning and AI, the document made three observations. Screening companies increasingly advertise these technologies but disclose little about what they do or what safeguards exist. Complex models can make the precise reason for a denial hard to identify and to disclose. And removing protected characteristics from a model's inputs does not settle the matter, because a machine learning system may re-weight other factors in a way that recreates a discriminatory outcome.

The guidance then set out six guiding principles under these headings:

1. Choose relevant screening criteria.
2. Use only accurate records.
3. Follow the applicable screening policy.
4. Be transparent with applicants.
5. Allow applicants to challenge negative information.
6. Design and test complex models for fair housing compliance.

It worked through the three kinds of record that screening reports usually carry. On credit, it said credit scores are built to assess the risk of default on a loan, not the risk of unpaid rent, that HUD was unaware of studies showing credit reports and scores accurately predict a successful tenancy, and that no HUD programme requires rental applicants to be screened on credit scores. It cited Urban Institute figures for August 2021 giving median FICO scores of 612 for Native American, 627 for Black, 667 for Hispanic and 727 for White individuals.

On eviction history, it said records should not be relied on when they are old, incomplete or irrelevant, and that proceedings the tenant won, settled or saw dropped should not support a denial. It cited a 2021 study of more than 3.6 million eviction records from 12 states which found 22 per cent ambiguous or falsely representing the tenant's history. On criminal records, it described as overbroad any policy that does not distinguish offences by nature, severity or age, that counts arrests without convictions, or that gives no opportunity to show mitigating factors. The guidance set no fixed look-back period for any of the three; it gave one HUD example, a policy found overbroad after a tenant was evicted over a forgery conviction 15 years old.

On responsibility, the guidance spoke to both sides of the transaction. Housing providers remained responsible for discriminatory outcomes when they outsourced screening, and were told to make their own decision on any denial recommendation and to tailor criteria instead of taking a product as supplied. Screening companies could be liable for practices that facilitate discrimination even where the provider makes the final decision. Disclaimers, the document said, generally cannot shield a party from liability under the Act.

Finally, it described what a transparent denial looks like: a letter stating each standard the applicant did not meet and how, with the records relied on attached and an explanation of how to appeal.

What the advertising guidance said

The second document addressed advertisers, meaning anyone placing adverts for housing, credit and real estate-related products such as rentals, property management, residential real estate services, mortgages and home insurance, and ad platforms, meaning the systems that direct and deliver adverts in digital spaces.

It described four functions that carry risk. Audience categorisation tools segment users by attributes such as gender, age, income, location or interests, and let an advertiser include or exclude a segment. Custom audiences are built from uploaded lists or tracked behaviour. Mirror audiences, sometimes called lookalike audiences, use an algorithm to extend a source list to people with similar traits. Delivery systems then use machine learning to predict who within the chosen audience is most likely to respond, and combine that with bids in real-time auctions. Pricing is a fifth pressure point: a system that charges more to reach one group can price an advertiser out of reaching it.

The guidance tied these functions to section 100.75 of title 24, headed "Discriminatory advertisements, statements and notices". That section lists, among its examples of prohibited conduct, selecting media or locations for advertising that deny particular segments of the housing market information about housing, and refusing to publish housing adverts or charging different fees or terms for them because of a protected characteristic.

The examples in the document were concrete: excluding families with children from the audience of a rental advert; aiming high-cost loans only at consumers with limited English proficiency; showing one neighbourhood a rental advert with warnings about credit and criminal records while a similar advert elsewhere carries none; and building a mirror audience from a tenant list that is mostly White and childless. It also noted that limiting explicit targeting options does not prevent discrimination further down the line, in delivery.

Its recommendations were split by role. Advertisers were told to use platforms that manage the risk of discriminatory delivery and to obtain disclosures on how they do so, to follow platform instructions for identifying housing adverts, to examine the source and make-up of custom and mirror audience data, and to monitor campaign outcomes. Platforms received seven recommendations, among them running housing adverts through a separate process and interface, avoiding targeting options that describe or stand in for protected characteristics, regular end-to-end testing, adopting less discriminatory alternatives for algorithmic systems, and making sure delivery does not produce different charges by protected characteristic.

The document also referred to United States v. Meta Platforms, a case in the federal court for the Southern District of New York, where the complaint was filed on 21 June 2022 and a settlement agreement was filed on 27 June 2022; the guidance cites that agreement when it suggests platforms consider disabling custom and mirror audience functions for housing adverts. On liability it took a broad view: any entity that plays a substantial role in a discriminatory housing outcome can be liable under the Act, even if it does not provide the housing.

Where each document stands today

The official record is a notice in the Federal Register of 6 April 2026, at volume 91, page 17291, titled "Notification of Withdrawal of Fair Housing and Equal Opportunity Guidance Documents" and signed by Craig W. Trainor, Assistant Secretary for Fair Housing and Equal Opportunity. It gives 17 September 2025 as the effective date of withdrawal and names eight documents.

The digital advertising guidance of 29 April 2024 is one of the eight. The others include two documents on assistance animals, dated 2013 and 2020, a 2021 statement on special purpose credit programmes, 2007 guidance on limited English proficiency, and a memorandum of 10 June 2022 on implementing guidance about the use of criminal records.

The notice cites Executive Order 14192, "Unleashing Prosperity Through Deregulation", and Executive Order 14219. It describes guidance as non-binding and says the withdrawn documents "have been removed from active use and should not be relied upon as authoritative". It also says that actions which do not comply with the text of the Fair Housing Act remain subject to enforcement by the department, that a person may still bring a civil action within two years, and that HUD is continuing its review and will reissue any guidance it finds necessary.

The tenant screening guidance is the less clear of the two. Its title is not among the eight in the notice. The press release that announced it sits on HUD's archive site under a banner reading "Content Archived: February 3, 2025", and the document itself is held on the same archive site. HUD's current Fair Housing Act overview page does not mention tenant screening algorithms, AI or digital advertising; it refers to a notice withdrawing FHEO guidance documents without naming them on the page, and says earlier guidance had discouraged criminal background checks for prospective tenants.

The proposal to remove the effects regulation

The regulation behind both documents is itself under review. On 14 January 2026 HUD published a proposed rule in the Federal Register, at volume 91, page 1475, titled "HUD's Implementation of the Fair Housing Act's Disparate Impact Standard" and signed by Secretary Scott Turner. It proposes to delete the sentence in section 100.5(b) that says a discriminatory effect may establish unlawful discrimination without intent, and to remove and reserve subpart G, which contains section 100.500.

HUD's stated reasons are Executive Order 14281, "Restoring Equality of Opportunity and Meritocracy", of 23 April 2025, which directs agencies to review and consider repealing disparate impact regulations; the two deregulation orders cited in the withdrawal notice; and the Supreme Court's 2024 decision in Loper Bright Enterprises v. Raimondo, after which, HUD says, courts give no deference to an agency's reading of a statute. The proposal states that courts, not a federal agency, are the appropriate place to decide questions about disparate impact liability under the Act. It does not mention algorithms, AI, tenant screening or advertising.

The regulation has changed hands before. The proposal recounts a first rule published on 15 February 2013, a replacement published on 24 September 2020 that a federal district court in Massachusetts stayed before it took effect, and a reinstatement of the 2013 standard published on 31 March 2023.

The comment period ran for 30 days, from 14 January to 13 February 2026, shorter than the 60 days HUD describes as its usual policy. No final rule appears in the text of the regulation: the Electronic Code of Federal Regulations, marked up to date as of 7 October 2026, still shows section 100.500 as current, with its 2023 source citation, and not as reserved or removed.

A recommendation can be withdrawn by the agency that wrote it. The statute it explained stays where Congress put it.

Consumer reporting law runs alongside

A tenant screening report is also a consumer report, and that brings in a second federal statute, the Fair Credit Reporting Act. HUD's screening guidance said it did not cover those obligations, which belong to other agencies. Their pages were not part of the withdrawal.

The Federal Trade Commission's guidance for landlords, published on 2 June 2023, says tenant background check reports are consumer reports and that the category includes risk scores or recommendations based on criteria the landlord selects. A landlord may obtain a report on a person applying to rent or renewing a lease, and must certify to the reporting agency that it will be used only for housing purposes.

When a landlord takes an adverse action based even partly on a report, a notice is due. The Commission lists as adverse actions denying the application, requiring a co-signer, requiring a deposit or a larger deposit than another applicant would pay, and charging higher rent than another applicant. The notice, required by section 615(a) of the Fair Credit Reporting Act, must give three things:

1. The name, address and phone number of the agency that supplied the report.
2. A statement that the agency did not make the decision and cannot give the specific reasons for it.
3. Notice of the right to dispute the accuracy or completeness of the information, and to a free report from the agency if requested within 60 days.

It may be given in writing, electronically or orally; the Commission calls written notice the best practice. Where a credit score was used, the written or electronic notice must also give the score, its source, date and range, and the key factors that lowered it, in order of importance.

The Consumer Financial Protection Bureau's consumer page on rental denials, last reviewed on 28 May 2024, adds the timetable for disputes: the screening or credit reporting company generally has 30 days to investigate, 45 days in some cases, and some states set shorter deadlines.

A worked example shows the calendar, on a simple count of days and with dates chosen for illustration. A landlord gives an adverse action notice on 1 July 2026. Sixty days later is 30 August 2026, the last day to ask the reporting company for the free report. If the applicant files a dispute on 10 August 2026, 30 days later is 9 September 2026 and 45 days later is 24 September 2026. How a particular deadline is counted depends on the case and on any shorter state period.

Both agencies also note the overlap with fair housing law: each page says a landlord who refuses to rent to anyone with a criminal record may violate the Fair Housing Act.

What the pages do not settle

Several questions are left open by the sources read for this guide. The formal status of the tenant screening guidance is the first: it is archived and absent from the list of eight, and no page consulted states that it was withdrawn or that it remains HUD's view. The second is the fate of the January 2026 proposal, on which no final rule was found in the regulation's text as of 7 October 2026. The third is how courts will treat effect claims about algorithms if HUD's regulation is removed, a question the proposal expressly leaves to them.

The pages also say nothing of state and local law. The Consumer Financial Protection Bureau notes that some states set shorter dispute deadlines; its page does not describe them, and this guide does not either. For any particular screening product, advertising campaign or application, the answer depends on the facts and on the law of the place.

Kooky, from Shaka

Kooky edits Agents Estate and builds Shaka, the payment router he made for real estate professionals. One payment comes in, and every agent, agency and party in the deal receives their signed share on closing date.

About Kooky and Shaka →