Inclusive Communities kept disparate impact in 2015 but required a policy shown to cause the gap
On June 25, 2015, the Supreme Court held disparate-impact claims cognizable under the Fair Housing Act, but said a statistical disparity fails without a policy that causes it. For lenders using models, each policy’s purpose and effect became a data question.
Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.
Two-minute summaryThe dek and the first sentence of each section, quoted from the article
On June 25, 2015, the Supreme Court held disparate-impact claims cognizable under the Fair Housing Act, but said a statistical disparity fails without a policy that causes it. For lenders using models, each policy’s purpose and effect became a data question.
- The week The dispute began with a federal tax program.
- The policy: what causes a disparity The Court gave an example.
- The statistics: what disparity evidence can show In our reading, this framed what statistics can do in such a case.
- The justification: the valid interest and its record For a lender, the valid interest of a credit policy can be expressed in terms of risk, cost or a legal requirement.
- The alternative: less discriminatory options and remedies For a lender using models, an alternative is a different model, a different feature set or a different threshold that achieves similar performance with less disparity.
- How the ruling reached different organizations For depository institutions that originate mortgages, the ruling confirmed the claim was available under the Act and described its limits. [1] In our reading, the immediate task was documentary: an inventory of credit policies, the interest each serves and the analysis behind thresholds.
- Objections, and where the argument stops One objection is that the safeguards made claims harder, so lenders could relax.
- Checks for leaders whose decisions run through models These are our own suggestions, not requirements of the ruling or any rule.
- With hindsight HUD returned to its rule after the decision.
The week
On June 25, 2015, the Supreme Court decided Texas Department of Housing and Community Affairs v. Inclusive Communities Project, Inc., No. 13-1371. [1] The case had been argued on January 21, 2015. [1] The Court held that disparate-impact claims are cognizable under the Fair Housing Act. [1] It affirmed the judgment of the Court of Appeals for the Fifth Circuit and remanded the case. [1] Justice Kennedy delivered the opinion of the Court, joined by Justices Ginsburg, Breyer, Sotomayor and Kagan. [1] Justice Thomas and Justice Alito filed dissenting opinions; Chief Justice Roberts and Justices Scalia and Thomas joined Justice Alito’s. [1]
The dispute began with a federal tax program. The federal government provides low-income housing tax credits that designated state agencies distribute to developers. [1] Congress directed states to develop plans identifying selection criteria for distributing the credits. [1] In Texas, the Department of Housing and Community Affairs distributed them. [1]
The Inclusive Communities Project, a Texas nonprofit that helps low-income families obtain affordable housing, sued the Department. [1] It brought a disparate-impact claim under sections 804(a) and 805(a) of the Act. [1] It alleged that the Department had caused continued segregated housing patterns by allocating too many credits to housing in predominantly black inner-city areas and too few in predominantly white suburban neighborhoods. [1]
The district court relied on statistical evidence. [1] It found that from 1999 to 2008 the Department approved tax credits for 49.7 percent of proposed non-elderly units in areas that were 0 to 9.9 percent Caucasian. [1] It found the Department approved 37.4 percent of such units in areas that were 90 to 100 percent Caucasian. [1] It found that 92.29 percent of tax credit units in the city of Dallas were in census tracts with less than 50 percent Caucasian residents. [1] It placed the burden on the Department to prove there were no less discriminatory alternatives, and ruled for the plaintiff. [1]
While the appeal was pending, the Department of Housing and Urban Development issued a regulation. [1] It interpreted the Act to encompass disparate-impact liability and set a burden-shifting framework. [1] Under it, a plaintiff first had to show that a challenged practice “caused or predictably will cause a discriminatory effect.” [1] The defendant then had to prove the practice was necessary to achieve one or more substantial, legitimate, nondiscriminatory interests. [1] The plaintiff could still prevail by showing those interests could be served by another practice with a less discriminatory effect. [1]
The Fifth Circuit held that disparate-impact claims were cognizable, but reversed on the merits. [1] Relying on HUD’s regulation, it held the district court should not have put the burden on the Department to prove there were no less discriminatory alternatives. [1] The Department petitioned the Supreme Court on whether such claims were cognizable at all. [1]
The Court said yes, and then set limits. It said disparate-impact liability “has always been properly limited in key respects.” [1] It said a claim relying on a statistical disparity “must fail if the plaintiff cannot point to a defendant’s policy or policies causing that disparity.” [1] It said an important way to limit liability was to give housing authorities and private developers “leeway to state and explain the valid interest served by their policies.” [1] And it said remedies should concentrate on eliminating the offending practice, with courts striving to design race-neutral means. [1]
The Inclusive Communities decision is remembered for its holding: disparate-impact claims are available under the Fair Housing Act. [1] For a chief data officer, a model risk manager or a fair lending officer at a mortgage lender, the opinion’s limits mattered as much. The Court said a statistical disparity is not enough. [1] A plaintiff has to point to a policy that causes it. [1] And a defendant has room to explain the valid interest the policy serves. [1]
The case was about a state agency’s allocation of tax credits, not about lending. [1] But the Court described section 805(a) as prohibiting discriminatory practices in real-estate transactions. [1] In our reading, its reasoning applies to any practice that can be challenged under that section. In our reading, its limits translated into data questions for any lender whose credit, pricing or marketing decisions run through models and rules.
Our argument is that the ruling made disparate-impact exposure a matter of how a lender can describe and evidence each policy in its data pipeline. In our reading, the Court’s safeguards protected lenders only to the extent they could show which policy produced an outcome and why it existed. They also needed to show whether an alternative would do the same job with less disparity.
The people who can answer those questions include data and risk staff as well as lawyers. Which rule applied to an applicant is a question for the system owner. Why a threshold sits where it does is a question for the credit risk team. Whether an alternative model performs as well is a question for data scientists. In our reading, the ruling moved part of fair lending defense into their work, and into the records their systems keep.
We take the ruling in four parts:
- The policy: what counts as the thing that causes a disparity.
- The statistics: what disparity evidence can and cannot show.
- The justification: the valid interest and the record behind it.
- The alternative: less discriminatory options and the shape of remedies.
The parts follow the order of a disparate-impact claim and do not overlap. The first is the object of the challenge, the second the plaintiff’s proof, the third the defendant’s answer and the fourth what happens if the answer is not enough. After them we look at how the ruling reached different organizations, take the objections, and end with checks a data or compliance leader could run.
The policy: what causes a disparity
The Court’s requirement
The opinion said a disparate-impact claim that relies on a statistical disparity “must fail if the plaintiff cannot point to a defendant’s policy or policies causing that disparity.” [1] It said its causality requirement ensures that racial imbalance does not, without more, establish a prima facie case. [1] It said this protects defendants from being held liable for disparities they did not create. [1]
The Court gave an example. A plaintiff challenging a private developer’s decision to build in one location rather than another “will not easily be able to show this is a policy causing a disparate impact.” [1] It said such a one-time decision “may not be a policy at all.” [1] It added that causation may be hard to establish because of the multiple factors that go into investment decisions. [1]
Policies in a lending pipeline
The opinion did not discuss lending models. In our reading, its distinction between a policy and a one-time decision maps onto how a lender’s decisions are made. A credit score cutoff, a pricing grid, a debt-to-income limit, an overlay above an investor’s guidelines and a marketing audience rule are each a policy in this sense. They apply to every applicant who meets their conditions.
A model is harder to classify. It is one artifact, but it combines many variables, each of which acts like a rule. In our reading, a plaintiff challenging a model would need to point to the model, or to a feature within it, as the policy. A lender that cannot itself say which features drive which outcomes will struggle to answer such a claim.
Overlays, exceptions and discretion
Not every lending decision is automated. Pricing exceptions, manual underwriting overrides and broker compensation choices involve judgment. The Court’s example of a one-time decision that “may not be a policy at all” is relevant here. [1] In our reading, a single exception is closer to a one-time decision. A pattern of exceptions, permitted by a written rule or by the absence of one, looks more like a policy.
That distinction makes exception data important. A lender that records each exception, its reason and who approved it can show whether discretion was bounded. A lender that does not record them cannot show what its practice was, in either direction.
Marketing and channels
Policies also operate before an application exists. Where a lender places branches, which audiences its online advertising reaches and which brokers it accepts all shape the applicant pool. The opinion’s discussion of investment decisions about where to build housing touched a similar issue for developers. [1] It said causation may be hard to establish because of the multiple factors behind such decisions. [1] In our reading, marketing and channel rules are policies a lender can document, even where their effects are hard to isolate.
Why identification is a data task
Identifying the policy that produces a disparity requires tracing decisions back to rules. For a manual process, that means written procedures and exception logs. For an automated one, it means version-controlled models, documented features and the ability to reproduce a past decision. The Court’s requirement put weight on the plaintiff to make that link. [1] In our reading, it also gave a defendant a reason to be able to make or refute the link from its own records.
Third parties and discretion
The opinion also addressed discretion. It said that if the plaintiff could not show a causal connection, “for instance, because federal law substantially limits the Department’s discretion,” the case should be dismissed. [1] In our reading, the same logic reaches a lender that follows an investor’s or insurer’s requirement it cannot change. HUD later took a different view of a codified defense along these lines, which we cover in hindsight. [2]
Version history as evidence
Policies change. A cutoff moves, a model is retrained, an overlay is added after a bad year. A disparity measured over several years may span several versions of the same policy. In our reading, a plaintiff pointing to “the policy” and a lender answering need to agree which version was in force when. A change log that records each version, its effective date and its reason turns that dispute into a lookup.
What the policy part asks of a lender
The policy part turns fair lending into an inventory question. A lender needs a list of the rules and models that affect who is offered credit, on what terms and through which channels. It needs to know which of them it controls and which it inherits from investors, insurers or vendors. And it needs to be able to say, for any outcome, which rules applied.
The statistics: what disparity evidence can show
Disparity is necessary, not sufficient
The Court said disparate-impact liability has been limited to avoid serious constitutional questions that might arise “if such liability were imposed based solely on a showing of a statistical disparity.” [1] It said a plaintiff who fails to allege facts at the pleading stage or produce statistical evidence “demonstrating a causal connection” cannot make out a prima facie case. [1]
In our reading, this framed what statistics can do in such a case. They can show that outcomes differ across groups. They cannot, on their own, show why. The causal link to a specific policy has to come from the analysis, not from the size of the gap.
The statistics in this case
The district court’s evidence shows the kind of data at issue. It compared approval rates for proposed units across areas grouped by their racial composition, from 1999 to 2008. [1] It measured the share of tax credit units in Dallas located in census tracts with less than 50 percent Caucasian residents. [1] The Supreme Court did not decide whether this evidence was enough. It said the case, on remand, “may be seen simply as an attempt to second-guess which of two reasonable approaches a housing authority should follow” in allocating tax credits. [1]
Choosing the comparison
Every disparity statistic rests on a choice of comparison. The district court compared approval shares for proposed units across areas grouped by racial composition. [1] It also measured where approved units were located in one city. [1] Each choice of grouping, period and geography produced a different picture.
For a lender, the same choices arise. Comparing approval rates among applicants differs from comparing who applied at all. Comparing outcomes in one market differs from comparing them nationally. In our reading, the Court’s insistence on a causal link meant the comparison had to fit the policy being challenged. A disparity in who applied says little about an underwriting cutoff, and much about marketing.
The choice also has to be stable. A lender that measures a policy one way this year and another way next year cannot show a trend, and invites the charge that it picked the comparison that looked best. In our reading, a documented testing method, applied the same way each period and changed only with a recorded reason, is itself part of the evidence a lender can offer. It shows that the analysis was designed before the results were known.
What causal analysis looks like for a lender
The opinion did not prescribe a method. In our reading, the requirement that a plaintiff tie a disparity to a policy suggests what a lender’s own analysis would need to show. It would compare outcomes with and without the policy, or with the policy’s threshold moved. It would control for factors the lender does not set, such as the applicant pool. It would separate the effect of one rule from the effect of others in the same pipeline.
Each of these is a data engineering task as well as a statistical one. The lender needs decision-level records, the inputs to each decision and the rules in force at the time. Without those, it cannot test the effect of a single policy in isolation.
Proxies and protected characteristics
Disparate-impact analysis asks about effects on protected groups, whether or not a policy uses group membership. In our reading, that creates a data tension. To test whether a policy has a disparate effect, a lender needs some way to estimate group membership for its applicants. Where the data are absent, the analysis depends on estimates. The opinion did not address this. It did say that “mere awareness of race” in solving housing problems does not doom an effort at the outset. [1]
Data a lender may lack
Some data needed for the analysis may not exist in a lender’s systems. Declined applications may be purged under retention schedules. Older model versions may not be reproducible. Vendor scores may arrive without the inputs behind them. In our reading, each gap limits what the lender can show about its own policies, and so limits the protection the Court’s safeguards offer.
Small numbers
Statistics need enough observations to mean anything. A small lender, or a small product line, may have too few applications from a group to measure a disparity reliably. In our reading, that cuts both ways. A plaintiff may struggle to show a disparity from small numbers. A lender may struggle to show the absence of one. Pooling data across periods or products, with the method recorded, is one way to test policies that small samples cannot.
Pleading and timing
The Court said courts must examine with care whether a plaintiff has made out a prima facie case and that “prompt resolution of these cases is important.” [1] In our reading, early resolution favors a defendant that can quickly show which policy was in force and what it did. A lender that needs months to reconstruct its own decisions loses the benefit of a fast test.
The justification: the valid interest and its record
Leeway to explain
The Court said an important way to limit disparate-impact liability is to give housing authorities and private developers “leeway to state and explain the valid interest served by their policies.” [1] It said this analysis is analogous to Title VII’s business necessity standard, which the Court had described as a defense in employment cases. [1] It said disparate-impact liability mandates the “removal of artificial, arbitrary, and unnecessary barriers,” not the displacement of valid governmental policies. [1]
The opinion described HUD’s regulation as requiring the defendant to prove the practice was necessary to achieve a substantial, legitimate, nondiscriminatory interest. [1] It quoted HUD’s statement that this step was analogous to the Title VII requirement that an employer’s interest be job related. [1]
The record behind a model
For a lender, the valid interest of a credit policy can be expressed in terms of risk, cost or a legal requirement. In our reading, the leeway the Court described is only usable if the lender can show the interest with evidence. For a model, that means documentation of what it predicts, how well it predicts it and why each feature is included. For a cutoff, it means the analysis that set the threshold.
Model risk management already asks for some of this. Fair lending asks a narrower question: does the documentation explain why this policy, with its effect on protected groups, serves the stated interest? In our reading, a model inventory that records purpose and performance but not group effects answers only half.
Public interests
The Court applied the same idea to government. It said that before rejecting a business justification, “or, in the case of a governmental entity, an analogous public interest,” a court must find that the plaintiff has shown an available alternative. [1] It said the Act is not an instrument to force housing authorities to reorder their priorities. [1] In our reading, a public agency running a lending or subsidy program needs the same record as a lender: what each criterion is for, and evidence that it serves that purpose.
Documentation at adoption and at validation
Two moments matter for the record. At adoption, the lender decides why a policy exists and what it is meant to achieve. At validation, it tests whether the policy does that. In our reading, a fair lending record needs both. A model validated for accuracy but never assessed for group effects leaves the justification untested. A model assessed for group effects but with no recorded purpose leaves the interest unstated.
Who approves a policy
Governance decides whether the record exists. A credit policy committee, a model risk committee and a fair lending function each see part of the picture. In our reading, the ruling gave lenders a reason to make fair lending review a step in approving a new model or threshold, not a later audit. The opinion does not say this. It follows from the leeway the Court described, which depends on what the lender can state and explain. [1]
Explaining a feature
For a model, the interest has to be explained at the level a challenge targets. If a plaintiff points to one feature, the lender needs to say why that feature is in the model and what it adds. In our reading, aggregate model accuracy does not answer that question. A record of each feature’s contribution, and of tests with the feature removed, does.
This is where model documentation and fair lending documentation meet. The model developer knows why a feature was added. The fair lending function needs that reason written down, with its effect on protected groups beside it.
Interests a lender cannot name after the fact
A justification assembled after a challenge is weaker than one recorded when the policy was adopted. The opinion did not say this. In our reading, it follows from the Court’s description of the defense as stating and explaining the interest a policy serves. A policy whose purpose was never written down invites the inference that it had none.
Profit and practical business choices
The Court said disparate-impact liability must be limited so regulated entities can make “the practical business choices and profit-related decisions that sustain a vibrant and dynamic free-enterprise system.” [1] It said that before rejecting a business justification, a court must find that the plaintiff has shown an available alternative practice with less disparate impact that serves the entity’s legitimate needs. [1] In our reading, this gave a lender room to defend a policy on commercial grounds, provided the commercial interest was real and documented.
The alternative: less discriminatory options and remedies
The third step
HUD’s regulation, as the opinion described it, let a plaintiff prevail by proving that the defendant’s interests “could be served by another practice that has a less discriminatory effect.” [1] The Court described the same idea from Title VII: before rejecting a business justification, a court must determine that the plaintiff has shown an available alternative practice. [1]
Searching for alternatives
For a lender using models, an alternative is a different model, a different feature set or a different threshold that achieves similar performance with less disparity. In our reading, a lender that has searched for such alternatives, and recorded the results, is better placed than one that has not. If a plaintiff later proposes an alternative, the lender can show whether it was tested and why it was rejected.
The search is a data and computing task. It needs candidate models, a performance measure, a disparity measure and a decision rule for trading them off. The opinion sets no such measures. In our reading, the third step made them worth defining in advance.
Trade-offs that need a decision rule
An alternative may not match the original on every measure. A model with less disparity may predict default slightly less well. A lower threshold may admit more applicants and more losses. In our reading, a lender searching for alternatives needs a rule, set in advance, for how much performance it will trade for how much reduction in disparity. Without one, each decision is made case by case, and is harder to defend as a policy.
The Court said a plaintiff must show an alternative that “serves the [entity’s] legitimate needs,” quoting an earlier case. [1] In our reading, that wording gave a lender room to reject alternatives that do not meet its needs, if it can show what those needs are and how the alternative fell short.
Remedies
The Court said remedial orders “should concentrate on the elimination of the offending practice.” [1] It said courts should strive to design additional measures to eliminate disparities through race-neutral means. [1] It said remedial orders that impose racial targets or quotas might raise more difficult constitutional questions. [1]
For a lender, the remedy language points to the policy, not the outcome. In our reading, a remedy would change the rule, model or threshold found to be the offending practice. A lender that can isolate the policy can also change it without rebuilding its whole pipeline.
Inherited requirements
Some policies a lender cannot change on its own. Investor guidelines, insurer requirements and program rules set by agencies apply to the loans a lender makes for them. The Court said that where federal law substantially limits a defendant’s discretion, a plaintiff may be unable to show the causal connection. [1] In our reading, a lender should know which of its policies are inherited, from whom, and how much discretion it keeps in applying them. Overlays added above an inherited rule are the lender’s own.
Monitoring after a change
A changed policy is a new policy. In our reading, a lender that replaces a model or threshold to reduce disparity needs to monitor the result, because effects in production may differ from effects in testing. The same records that support the original justification support the change: what was replaced, why, and what happened next.
Quotas as a risk the Court named
The Court also explained why it limited the doctrine. Without adequate safeguards, it said, potential defendants may adopt racial quotas, which itself raises serious constitutional concerns. [1] It said courts should avoid interpreting disparate-impact liability so broadly as to inject racial considerations into every housing decision. [1] In our reading, the opinion asked lenders to manage disparate effects through the design of policies, not through targets for outcomes.
How the ruling reached different organizations
Banks and credit unions
For depository institutions that originate mortgages, the ruling confirmed the claim was available under the Act and described its limits. [1] In our reading, the immediate task was documentary: an inventory of credit policies, the interest each serves and the analysis behind thresholds. Credit unions with smaller data teams would face the same questions with fewer resources, and may rely more on vendors’ documentation.
Nonbank lenders and fintechs
HUD’s 2026 proposal summarized the Act as prohibiting discrimination in the sale, rental or financing of dwellings and in other housing-related activities. [4] In our reading, that reaches nonbank lenders, including those using automated underwriting or alternative data. In our reading, firms that retrained models at short intervals had the hardest version of the policy problem. Each model version is, in effect, a new policy, and the record of which version made which decision becomes evidence.
Insurers
HUD’s 2023 rule recounted its 2016 statement that categorical exemptions or safe harbors for insurance practices are unworkable. [2] Its 2023 rule cited a case allowing allegations that an insurer’s credit scoring system had an unjustified discriminatory effect. [2] In our reading, insurers using predictive models faced the same data questions as lenders.
State housing finance agencies
The case itself was against a state agency allocating federal tax credits. [1] The Court said it did not impugn housing authorities’ race-neutral efforts to revitalize communities, and that the Act is not an instrument to force authorities to reorder their priorities. [1] For a state housing finance agency, the ruling meant its selection criteria could be challenged, and that its stated interests would get leeway if explained. In our reading, the agency’s scoring criteria were its policies, in the same sense as a lender’s cutoffs.
Appraisers and valuation
The opinion discussed appraisal directly. It described a 1988 amendment providing that nothing in the Act prohibits a person in the business of furnishing appraisals from considering factors other than race, color, religion, national origin, sex, handicap or familial status. [1] The Court read that exemption as a sign that Congress recognized disparate-impact liability under section 805(a). [1]
In our reading, valuation sits in the same position as credit. A valuation model or appraisal practice can be a policy with group effects. The 1988 provision permits considering non-protected factors. It does not, on the Court’s reading, place valuation outside the Act.
Public lending and assistance programs
State and local agencies also run lending and assistance programs, such as down payment help and rehabilitation loans. Their eligibility rules, scoring criteria and partner lists are policies in the Court’s sense. [1] In our reading, an agency that buys eligibility or scoring tools from vendors faces the same documentation questions as a lender. The Court’s reference to an analogous public interest gives it similar leeway if it can explain its criteria. [1]
Vendors of models and data
Lenders can buy models, scores and data from vendors. The ruling did not address vendors. [1] In our reading, a lender that cannot see inside a vendor’s model cannot identify which feature drives a disparity, or explain its interest. Contracts that give the lender documentation, validation results and the right to test for disparate effects became part of fair lending readiness.
Objections, and where the argument stops
“The ruling helped defendants, so data work is less urgent”
One objection is that the safeguards made claims harder, so lenders could relax. The Court did describe its limits as protecting defendants. [1] But each safeguard works only if the defendant can use it. Pointing out that a plaintiff has not identified a policy is easier if the lender knows its own policies. Explaining a valid interest is easier if the interest is recorded. In our reading, the safeguards rewarded preparation.
“The case was about tax credits, not lending”
A second objection is that the facts were far from lending. They were. [1] The holding, though, concerned the Act’s sections 804(a) and 805(a), and the Court described section 805(a) as a prohibition of discriminatory practices in real-estate transactions. [1] It noted that a 1988 exemption for real-estate appraisers was in the same section. [1] It read that as Congress recognizing disparate-impact liability under that section. [1]
“Federal policy turned against disparate impact”
A third objection is that later federal policy made the ruling less relevant. HUD and the White House did move against disparate-impact liability from 2025, as we describe in hindsight. [3] [4] But the Court’s holding is about the statute, not about agency rules. HUD’s 2026 proposal itself said it would leave questions of interpretation “to courts.” [4] In our reading, private plaintiffs can still bring claims under the Act, so the data questions remain.
“Testing for disparity invites the quotas the Court warned against”
A fourth objection is that measuring group outcomes pushes lenders toward managing to targets, which the Court said raises constitutional concerns. [1] The Court did warn that without safeguards, defendants may adopt racial quotas. [1] It also said that “mere awareness of race” in solving housing problems does not doom an effort. [1]
In our reading, the answer lies in what a lender does with the measurement. Testing whether a policy has a disparate effect, and searching for a race-neutral alternative, matches the Court’s description of remedies that concentrate on the offending practice. [1] Setting outcome targets by group does not. The distinction is between changing policies and managing outcomes.
“Models are too complex to point to a policy”
A fifth objection is that modern models cannot be broken into policies, so the Court’s requirement protects any lender that uses one. In our reading, that is a risky assumption. A plaintiff can point to the model itself as the policy. The lender then needs to explain the interest it serves and whether an alternative would do. HUD’s 2023 rule said its framework applies to cases involving algorithms and machine learning. [2] Complexity makes the lender’s explanation harder to give, not unnecessary.
Where the argument stops
We do not say how any court has applied the ruling to a lending model. We do not cover the Equal Credit Opportunity Act, which has its own regulators and history, except where HUD and the executive order mention it. [3] Nor do we predict how courts will treat disparate-impact claims if HUD’s regulation is removed. We do not assess any lender’s practices. The opinion does not mention credit models, and our application of it to them is our own reading. Organizations should take their own legal advice on the Act’s application to their products.
Checks for leaders whose decisions run through models
These are our own suggestions, not requirements of the ruling or any rule.
Inventory
- List every rule and model that affects who receives credit, on what terms and through which channel.
- For each, record who owns it, whether it is internal, investor-required or vendor-supplied, and when it changed.
- Keep decision-level records that let you reproduce which rules applied to a past applicant.
Purpose and effect
- Write down the interest each policy serves when it is adopted, with the analysis behind any threshold.
- Test each policy’s effect on protected groups, using a documented method for estimating group membership where data are absent.
- Record searches for less discriminatory alternatives, including the options rejected and why.
Oversight
- Make fair lending review a step in approving any new model, threshold or overlay.
- Report to the responsible committee which policies were tested for group effects, and which were not.
Vendors and change
- Require vendors to provide documentation, validation results and the ability to test for disparate effects.
- Treat each model version as a policy change, with its own purpose, test results and approval.
- Track the status of federal rules separately from the statute, since the agency position has changed several times.
With hindsight
HUD returned to its rule after the decision. In 2018 it issued an advance notice asking what changes the ruling required, and in 2019 it proposed amendments. [4] On September 24, 2020, it published a final rule that it said implemented the Court’s decision. [4] HUD’s 2023 rule later described that 2020 rule as changing pleading standards, altering the burden-shifting framework and adding defenses. [2]
One of the added defenses concerned models. HUD’s 2023 rule said the 2020 rule created an “outcome prediction” defense, which commenters said was designed to shield certain policies based on algorithms. [2] The 2023 rule recounted HUD’s view, stated in its proposal, that the defense would in practice exempt most insurance practices and many other practices relying on outcome predictions, such as lending, from disparate-impact liability. [2]
The 2020 rule never took effect. HUD’s 2023 rule recounted that the U.S. District Court for the District of Massachusetts issued a preliminary injunction staying it before its effective date. [2] It cited that court’s October 25, 2020 decision, which found the outcome prediction defense and other changes likely unsupported by the Supreme Court’s ruling or other decisions. [2] The 2013 rule therefore remained in effect. [2]
On March 31, 2023, HUD published a final rule reinstating the 2013 rule and rescinding the 2020 rule, effective May 1, 2023. [2] It said the outcome prediction defense was unclear and not found in any case law. [2] It said the defense could improperly exempt practices “increasingly reliant upon algorithms and automated processes that rely on outcome predictions, such as lending practices.” [2] It said its rule described the framework to be used in cases involving algorithms and machine learning. [2]
The federal position then changed again. Executive Order 14281 is dated April 23, 2025 and was published on April 28, 2025. [3] It set a policy of eliminating the use of disparate-impact liability in all contexts to the maximum degree possible. [3] It directed agencies to deprioritize enforcement of statutes and regulations to the extent they include such liability. [3] It directed HUD, the CFPB and other agencies enforcing the Equal Credit Opportunity Act and the Fair Housing Act to evaluate, within 45 days, pending proceedings relying on it. [3]
On January 14, 2026, HUD proposed to remove its discriminatory effects regulations. [4] It said it would leave questions about interpreting disparate-impact liability under the Act to the courts. [4] It cited the executive order and the Supreme Court’s Loper Bright decision on agency deference. [4] On August 10, 2026, HUD published a supplemental proposal that also proposed removing disparate-impact provisions from its Title VI regulations. [5] It reopened comment on the January proposal for related topics, with comments due by October 9, 2026. [5] As checked on November 16, 2026, the Federal Register listed no final rule on either proposal, and the August notice describes the removal as proposed, not final.
What held up from 2015 is the Court’s holding. HUD’s 2026 proposal recounted that the Court held disparate-impact claims cognizable under the Act. [4] Removing a regulation does not change a statute the Court has interpreted. [1] [4] What did not hold up is any expectation that the federal agency rule would settle how the doctrine applies to models. Across 2020, 2023 and 2026, HUD wrote, withdrew and then proposed to remove the rules that addressed algorithms most directly. [2] [4]
Spotted an error? Tell us and we will check it and correct the article where needed.