Regulation B(eware): Is Algorithmic Debiasing Now Intentional Proxy Discrimination?

The Regulation B changes were expected, but—for some—they have created severe disruption.

On April 22, 2026, the Consumer Financial Protection Bureau ("Bureau") finalized a rule to reverse the Federal Reserve Board's ("FRB's") long-standing interpretation that the Equal Credit Opportunity Act ("ECOA") prohibits disparate impact discrimination.[1] Effective July 21, 2026, every reference to the FRB's “effects test” language disappears from Regulation B, and the official interpretation is that ECOA is a disparate-treatment-only fair lending statute. This is a genuinely momentous change, and it has been the lede of virtually every news story about the event.

But for many non-mortgage lenders, this is not the part that should keep them up at night. The change that matters more for them is quieter — a simple clause, but one with substantial force. It reflects, likely intentionally, the Supreme Court's current majority thinking that race-conscious decision-making—regardless of which group it favors — violates federal anti-discrimination laws, even though such laws were enacted primarily to protect certain historically-disadvantaged groups.

"The Act does not provide for the prohibition of practices that are facially neutral as to prohibited bases, except to the extent that facially neutral criteria function as proxies for protected characteristics designed or applied with the intention of advantaging or disadvantaging individuals based on protected characteristics." (emphasis mine) - Official Interpretations §1002.6(a)(2)

Read that twice. The new standard (like the old) still finds the intentional use of proxies to be a form of disparate treatment; however, it explicitly states that the direction a proxy effect runs is irrelevant. A facially-neutral variable wielded to help a protected class is captured by the same language as one wielded to hurt it. And this is particularly relevant now as it captures precisely the mechanism at the heart of modern, algorithmically-debiased, less discriminatory alternative ("LDA") credit models. Let's see why.

The Emergence of Modern-Day LDA Credit Models

To understand how this Regulation B ("Reg B") change ties into modern-day LDA credit models, let's go back about five years to when a cohort of regulatory technologists rose to prominence in fintech lending, regtech consulting firms, and inside the Chopra-era Bureau. With a powerful toolbox of machine learning algorithms, these talented professionals set out to re-engineer modern-day credit risk modeling through an equitable lending lens. However, whether by intention or ignorance, the important guardrails the Supreme Court ("SCOTUS") had thoughtfully laid out in Texas Department of Housing and Community Affairs v. Inclusive Communities Project ("Inclusive Communities") to keep disparate impact from collapsing into racial balancing were absent.[2]

Under those guardrails, the Court held that disparate impact was cognizable under the Fair Housing Act ("FHA"), but fenced it in carefully, precisely to avoid pushing lenders toward race-conscious, quota-like behavior:

“Without adequate safeguards at the prima facie stage, disparate-impact liability might cause race to be used and considered in a pervasive way and ‘would almost inexorably lead’ … to use ‘numerical quotas,’ and serious constitutional questions then could arise.” — Inclusive Communities

Accordingly, SCOTUS required a disparate impact claim to identify a specific policy as the cause of the disparity, and recognized liability if that policy created an “artificial, arbitrary, and unnecessary” barrier to credit. When these conditions held, the remedy was similarly limited to the elimination or repair of the offending factor:

“A disparate-impact claim relying on a statistical disparity must fail if the plaintiff cannot point to a defendant’s policy … causing that disparity.” … “Remedial orders … should concentrate on the elimination of the offending practice.” — Inclusive Communities

However, as I detailed in my Algorithmic Justice article, under the technologists' new equitable lending focused framework, none of the Inclusive Communities requirements appear present in their modern-day LDA credit models. With the click of a mouse, most credit scoring models are deemed discriminatory based on the summary judgment of an outcomes-based AIR-judge. Such models are immediately sent for algorithmic rehabilitation using automated black box debiasing software provided by a growing number of fintech and regtech startups. These debiasing programs treat the model as a whole as the offender,[3] render the business-necessity inquiry a tautology, and then hand the entire model over to a "debiasing" algorithm that uses demographic data—potentially in violation of ECOA—to adjust the weights of the model, or to alter the composition of the model's predictive factors. The explicit goal of this process (i.e., its intent)? To achieve a more balanced set of demographic approval rates at the expense of an "acceptable" trade-off in the model's predictive accuracy.

No specific, causal credit factor is identified as the cause of the AIR-based disparity.[4]  No credit factor is tested against SCOTUS's “artificial, arbitrary, and unnecessary” standard. The remediation is not targeted at the offending factor; instead, it is a model-wide rebalancing tuned to a protected-class-defined outcome. And whether the alternative model is truly less discriminatory is based on a narrow assessment of the model's decision impacts. While these LDA credit models may certainly be less discriminatory (as measured by approval rates) for one side of the protected-class line, they are decidedly more discriminatory for the other — a clear case of reverse disparate impact used as a remediation tool, as I have analyzed extensively in my Fool’s Gold series of articles.[5]

The Risks of AIR-Based Remediation

Under the Bureau's revised Regulation B, this sustained departure from Inclusive Communities permitted by (and, in some cases, encouraged by) past Bureau and DOJ leadership is now being called out for what it appears to be—disparate treatment by intentional use of a demographic proxy. And more recent developments indicate that I am not alone in this view. For example, attorneys from a leading consumer compliance law firm practice stated the following on a recent podcast:[6]

"... I think the CFPB did still warn creditors at some point in the rulemaking release that sort of an explicit racial balancing of a model to equalize adverse impact ratios or something like that would still be viewed as illegal in their view, that that would be disparate treatment in their consideration."

"Right. And I think there's a good balance there between testing to see how your model is performing and then reacting to it. So reacting to it and making adjustments based on what you see, that's where you have to be careful and make sure you're staying away from any balancing or anything that could really be seen as disparate treatment."

Perhaps even more telling, on May 27, 2026, the National Fair Housing Alliance and other plaintiffs, including a leading regtech consulting firm, filed a lawsuit challenging these Reg B changes. The standing of the regtech firm to file suit? By its own admission, these changes are adversely affecting the demand for plaintiffs' algorithmic debiasing services and harming its financial condition:

"The CFPB’s Final Rule also diminishes demand for [plaintiff's] services because it expresses a view that mitigating disparate-impact risks encourages or requires disparate treatment. Through this assertion, the CFPB has signaled that it is likely to take the position that analyses like those offered by [plaintiff] designed to mitigate disparate-impact risks are, themselves, illegal."

"[Plaintiff] has suffered economic and reputational harm as a direct result of the Final Rule. Clients have reduced the resources they expend on [Plaintiff's] services, while other entities have paused the engagement agreement process in anticipation of the Final Rule. [Plaintiff] expects other clients and entities to continue to reduce resources expended on [Plaintiff's] services. [Plaintiff] was also forced to delay a capital raise based on its disparate-impact product because the CFPB’s changes to Regulation B affected the market demand for [Plaintiff’s] disparate-impact product and services." - NFHA, et. al. vs. CFPB, et. al. (May 27, 2026)

Under the previous Reg B interpretation (and prior Bureau leadership), a lender deploying an LDA credit model could say: “We tested our model for disparate impact, found an AIR concern, and adopted a less discriminatory alternative to improve approval rate equity.” Under the new Regulation B and federal bank regulatory regime, the same facts read very differently: “We took a facially neutral model, used protected-class data to re-tune it with the express objective of changing protected-class outcomes, and adopted the result.”

The first sentence describes fair lending compliance risk management under a discrimination theory the Bureau has just disclaimed. The second describes the defining example of an intentionally applied proxy.  The facts did not change. The doctrine did. Whether a given LDA credit model actually crosses the line is, in the end, a legal determination that turns on how it was built, what was documented, and how courts and regulators read the new text — and none of that is settled. But as a matter of compliance risk, the shift is unfavorable.

In fact, three features of typical LDA credit model implementations push this out of a simple compliance manual update and onto the agenda of a lender's risk management committee.

• The documentation trail is usually problematic. Unlike some business decisions where intent can be hard to prove, LDA credit models arrive with explicit technical write-ups that name the fairness objective, identify the protected classes used in tuning, and quantify the AIR improvement achieved. A plaintiff, federal regulator, or Attorney General would not have to infer intent. The project plan and documentation typically states it for them.

• Prior regulatory encouragement is not a safe harbor. While it is true that prior Bureau leadership encouraged LDA credit model search and the use of automated debiasing methods as documented in the since-rescinded Supervisory Highlights Advanced Technologies Special Edition issued in January 2025, that encouragement was tethered to a regulatory theory that current Bureau leadership has now repudiated. A defense built on the old agency posture is plausible, but likely fragile.

• Multi-jurisdictional disparate impact postures complicate the path forward. A national lender now faces a federal regulator that may view protected-class-driven tuning as disparate treatment, while several state attorneys general may view the failure to perform disparate impact testing and remediation as a state law violation. Those theories are not strictly contradictory — a lender can document business-necessity standards, run causal-factor analyses, and avoid intent-laden, proxy discrimination-based remediation all at once — but threading that needle takes a more deliberate disparate impact program design than most LDA credit model workflows currently embody.[7]

Before turning to how Compliance Officers may wish to respond to these Reg B changes, I want to first address some of the counterarguments I have read as to why this final rule is misguided, and why the Bureau's claims about intentional proxy discrimination are, in fact, not applicable to modern-day LDA credit models.

Evaluating The Technologists’ Defenses

In the Bureau's commentary associated with the Final Rule, as well as in the lawsuit noted above, there are several counterarguments put forth as to why these Regulation B changes are misguided, unnecessary, and inapplicable to modern-day LDA credit models. I will not address all of them, but focus on those that I view as most important. Once again, I am focusing only on algorithmically-based LDA credit models.

1. LDA credit models do not treat similarly-situated credit applicants differently. There is therefore no disparate treatment.

This is generally true of LDA credit models based on the traditional remediation of "artificial, arbitrary, and unnecessary" causal factors, but frequently false for LDA credit models based on AIR-based algorithmic debiasing. In the latter case, when a debiasing algorithm re-weights facially-neutral variables to improve the AIR, two otherwise-identical applicants who differ only in a feature(s) correlated with their protected classes can receive different scores and different credit outcomes, relative to the original credit model, precisely because that feature was selected for its demographic signal. The claim that no one is treated differently holds only if you define “differently” to exclude differences introduced through a protected-class-tuned variable(s)—which is the very thing in dispute. The mechanism does not avoid disparate treatment; it relocates it from the protected class attribute itself to an algorithmically-determined proxy derived from facially-neutral predictive factors.

2. Credit is not a zero-sum game.

As a statement pertaining to the removal of an “artificial, arbitrary, and unnecessary” credit factor, this is correct. An irrelevant factor can skew credit outcomes of otherwise creditworthy individuals (e.g., removing a minimum loan amount of $50,000 from otherwise creditworthy individuals increases loan approvals). But a model optimized jointly for predictive accuracy and an AIR penalty is not removing a barrier; it is reallocating approvals and denials within a given credit policy threshold to increase approval rate equity in exchange for a certain amount of predictive accuracy (i.e., approving certain non-creditworthy applicants and denying other creditworthy applicants).

For example, if the lender's credit policy indicates that new originations, however aggregated (by campaign, by month, etc.), should not exceed an expected default or loss rate of X%, or should generate a measure of expected financial return of no less than Y%, then these credit policies create a zero sum game at the margin — and not only across demographic groups but within them as well. As my prior Fool's Gold analyses have shown empirically, LDA credit models expand protected class approvals (and reduce control group approvals) to stay within the lender's overall credit policy thresholds.[8] However, this reallocation also creates material credit decision changes within each demographic group since the algorithmically-derived proxies are imperfect.[9] This results in some protected class (and control group) applicants being denied under the LDA credit model even though such applicants were approved under the original credit model.

Overall, there are clear winners and losers with AIR-based remediation, and some of the losers belong to the very group the remedy claims to help. That is clearly zero-sum at the margin, twice over. The “not zero-sum” framing is true of the traditional remedy aimed at the removal of irrelevant causal factors, but false for the AIR-based remedy typically deployed.

3. Proxies are not considered disparate treatment ... until they are.

In my opinion, this is the biggest flaw in the technologists' defenses as it rests on an irreconcilable logical inconsistency. On the one hand, an unintentional facially-neutral variable that disadvantages a traditional protected class group is condemned as a proxy whose disparate treatment must be mitigated. This has been a concern about credit scoring models for decades, and was most recently highlighted as a risk by the Bureau in its since-rescinded Supervisory Highlights Advanced Technologies Special Edition issued in January 2025. On the other hand, the very same kind of facially-neutral variable, used intentionally to advantage a traditional protected class group, is reframed as a justifiable method of disparate impact remediation that should be encouraged.

This is a tortured definition of demographic proxy whose legality hinges on its situational direction of effect. And both SCOTUS and the new Reg B do not tolerate it. Using an advantaging proxy as a discrimination cure appears to be completely at odds with Inclusive Communities as it requires remediation targeted to the specific offending policy, not a model-wide re-weighting that offsets one disparate impact by manufacturing an opposing one. Reg B now specifically addresses it for non-mortgage credit— its intent-based test does not care which way the proxy runs, reaching “advantaging or disadvantaging” alike.

How Should Compliance Officers Respond?

While the Final Rule takes effect July 21, 2026, its implementation could certainly be stayed by a Court, and any judicial resolution could be months or years away with an uncertain resolution. However, maintaining a compliance program that bets against implementation of the new rule may not be a risk worth taking—but that is a decision for a Compliance Officer to make with its senior management, board, and counsel. So, without crossing into legal advice, here are some questions you may wish to consider now to evaluate a potential path forward.

• For each LDA credit model in production, do the development artifacts state, in substance, that protected-class data were used to re-tune weights, or to select the specific set of predictive factors, in the pursuit of greater protected class approvals or a more balanced AIR (or other similar comparative outcome metric)?

If the answer is yes, intent would appear to be documented, and the federal disparate-treatment exposure would appear non-trivial. It may be prudent to explore other more traditional disparate impact testing and remediation approaches (see below).

• Are any of these models used for housing-related credit, or for credit originated in states that still enforce an effects test? If yes, the model cannot simply be retired: Inclusive Communities and HUD’s effects-test rule still govern housing credit, and those state regimes preserve a disparate impact theory, so pulling the model may likely re-introduce exposure under the FHA or state law.

In these cases, as above, you may wish to explore other more traditional disparate impact testing and remediation approaches that do not intentionally focus on outcome re-balancing across demographic groups—such as those described in my Algorithmic Justice article.

• Can the lender cleanly distinguish causal-factor remediation (defensible across both regimes) from outcome-balancing remediation (now squarely in tension with the federal rule)?

The former is far easier to defend; the latter is the exposure.

Final Thoughts

For over five years, “algorithmic debiasing” became the focus of fair lending disparate impact compliance—a button to press when an AIR test failed, and a remedy whose mere existence was said to prove the original model could not survive Step 3 of the burden-shifting framework (i.e., a less discriminatory alternative was highly likely to be found). I was skeptical of its grounding under Inclusive Communities. Under the revised Regulation B, I am considerably more so.

The new federal framework does not punish lenders for caring about fair outcomes. It punishes them for producing those outcomes by intentionally using a facially-neutral variable as a proxy for a protected characteristic. That is not hyperbole. On a fair reading of how most LDA credit models are actually built, it is a reasonably literal description of the procedure—and lenders relying on those models, particularly outside the housing context and in jurisdictions where federal ECOA is the operative regime, should be reviewing them, with counsel, post haste.

The alternative is the one I have argued for from the start: ground disparate impact analysis in causal-factor identification, narrowly tailored remedies, and the substantive law—not in one-click algorithmic rehabilitation processes designed by those whose focus on elegant algorithms and equitable outcomes supersedes their apparent adherence to governing legal frameworks. That traditional approach is not perfect. But it is defensible under disparate impact where it still applies, and it does not carry the signature of intentional proxy use under the theory that has just replaced it.

Endnotes

[1]  Prior to the formation of the Bureau, the FRB had the authority to write and maintain ECOA's implementing regulation—Regulation B. In its original drafting, the FRB included certain "effects test" language in the regulation not based on the specific text of the ECOA, but on its review and interpretation of the Act's legislative history and the corresponding Congressional Record. This decision has been subject to debate and criticism since the regulation was issued, with many in the consumer lending industry and legal community eagerly awaiting an opportunity for the Supreme Court to weigh in on this interpretation.

[2] While Inclusive Communities specifically addressed disparate impact under the Fair Housing Act, most industry and legal practitioners leveraged the opinion to similarly interpret disparate impact requirements under the ECOA.

[3] Proponents focus on the scoring model as a whole for this treatment—rather than one or more specific causal factors—due to the sheer complexity of AI credit models, some containing hundreds or thousands of predictive factors. They argue that this complexity makes specific identification of the offending credit criteria impossible or impractical to determine. As I have noted elsewhere, this argument often conflicts with the proponents' positions on model explainability, where they cite the ability of ex post explainability methodologies, such as SHAP or integrated gradients, to provide reliable explanations for why the model made the specific decisions it did.

[4] There is an even deeper flaw in the technologists' LDA credit models relative to Inclusive Communities, one I documented extensively in my Fool’s Gold 4 article on LDA Credit Model instability. Specifically, the LDA credit models produced by these automated debiasing methods are inherently unstable: that is, a small change in the training sample, or a small adjustment to the fairness weight, can send the debiasing algorithm off to select an entirely different set of predictive factors to re-weight. This result evidences a strong logical break with the foundational basis of disparate impact laid out by SCOTUS. If the factors driving a disparity are genuinely “artificial, arbitrary, and unnecessary” barriers to credit—real causal culprits—their identification would not change in response to a trivial perturbation of the data. A remedy whose target moves that freely is not identifying an offending causal practice; it is chasing an outcome.

[5] This same type of algorithmic debiasing may also impact a lender's digital advertising models and algorithms in light of Meta's 2022 Fair Housing Act settlement with HUD and the DOJ. In particular, if the same type of digital advertising algorithms for non-mortgage credit products represent a potential violation of ECOA's prohibitions on discouragement, and if a lender mitigates this risk by adopting debiasing technologies similar to Meta's Variance Reduction System, then the Reg B changes highlighting intentional proxy discrimination would appear to similarly apply.

[6] CFPB’s Reg B Final Rule: Disparate Impact Liability Out, Discouragement Standard Narrowed, and SPCPs in the Crosshairs, The Consumer Finance Podcast, Troutman Pepper, May 21, 2026, at 06:54 and 07:12.

[7] For more on the traditional approach to disparate impact remediation for credit models, see Algorithmic Justice: What's Wrong With the Technologists' Credit Model Disparate Impact Framework.

[8] Importantly, as I have empirically shown in my Fool's Gold articles, this result for modern-day LDA credit models is a facade. The default probabilities estimated by these models, while accurate on average across the entire applicant pool, understate the expected defaults and losses on new LDA-approved loans. This occurs because the algorithm's factor re-weighting intentionally underestimates their absolute riskiness (but largely preserves their rank-ordering) to make them approvable. The underlying true credit risk of these new approvals hasn't changed—its just being intentionally suppressed to mechanistically boost the AIR.

[9] That is, model factors that are down-weighted (to increase the likelihood of approval) may be disproportionately concentrated in protected class applicants, but not completely so. Similarly, model factors that are up-weighted (to decrease the likelihood of approval) may be disproportionately concentrated in control group applicants, but not completely so. These imperfect demographic proxies cause a second-order reallocation of approvals and denials within each demographic group.

© Pace Analytics Consulting LLC, 2026.