The Two Theories Behind a Neutral Credit Rule
The practical problem in revised Regulation B is deciding what a neutral credit rule means when it produces different results for different groups. This chapter separates three possibilities—intentional disparate treatment, disparate impact based on effects alone, and a neutral rule used as an intentional proxy—so readers can understand what demographic evidence can still show and what it cannot establish by itself.
That distinction is easy to lose because all three paths can begin with the same observation: one group is approved less often, pays more, or receives fewer opportunities. The number is important, but it does not identify its own cause. The legal question is whether the protected characteristic was an intended reason, merely an unequal consequence of a neutral rule, or evidence that the rule was designed or applied to conceal an intentional difference.
Three paths from a credit rule to an applicant
Disparate treatment is the most direct path. It occurs when a creditor treats an applicant differently because of a protected characteristic. The difference may be explicit, as when a policy names the characteristic. It may also be hidden behind a neutral explanation: a lender can describe a criterion in general terms while selecting, applying, or enforcing it because it stands in for race, sex, or another protected basis. In either case, the protected characteristic has a role in the decision-making reason.
Disparate impact begins with a facially neutral policy. “Facially neutral” means that the rule does not openly sort applicants by a protected characteristic and is formally applied in the same way. Yet the rule may have a disproportionately adverse effect on a protected group. Under an effects-test theory, that effect can be the legal center of the claim even when the creditor did not intend to discriminate. The analysis then focuses on the relationship between the policy, the disparity, and the creditor's justification under the governing statute.
The 2026 federal ECOA position targets that second path. The CFPB's amended commentary says the Act does not provide for a prohibition based only on the effects of a facially neutral criterion. The FTC policy statement likewise says the Commission will no longer pursue disparate-impact claims. This is a change in the theory that can independently carry the claim; it is not a claim that unequal outcomes never happen or never deserve investigation.
The third path is where the boundary becomes more demanding rather than empty. The CFPB says a facially neutral criterion can remain unlawful when it functions as a proxy for a protected characteristic and is designed or applied with the intention of advantaging or disadvantaging people on that basis. A proxy is not unlawful merely because it correlates with a protected group. The key additional question is why the creditor chose or used it, and whether the surrounding facts support an inference of intentional substitution.
Imagine a lender adds a neutral-looking neighborhood variable to an underwriting model. The approval rate for a protected group falls. That observation alone fits an effects story, but it does not tell us whether the variable is a sensible risk measure, an intentionally selected substitute for a protected characteristic, or one part of a broader decision process. Investigators would need to examine the variable's purpose, the process by which it entered the model, how staff use it, whether exceptions are granted consistently, and what communications or internal records reveal about the decision. The same outcome gap can therefore be the beginning of different inquiries.
The word “neutral” describes the surface of a rule, not necessarily its full causal history. A useful analysis separates at least four stages: why the creditor selected a criterion, how the criterion was translated into a policy or model, how decision-makers applied it to individual applicants, and what outcomes followed. A disparity located at the fourth stage does not by itself prove that the first three stages contained discriminatory intent. But it can direct attention toward them. If internal records show that a variable was chosen because it tracked a protected group, or that staff were told to use it differently for different applicants, the facts no longer describe only an unintended effect.
The same neighborhood variable can therefore produce different conclusions under different evidence. Suppose it was adopted after testing repayment risk, used consistently, and never connected to a protected classification in the lender's decision process. The observed gap may remain a serious business and policy concern, but under the revised federal ECOA position it is not, without more, the effects-test claim formerly described in Regulation B. Suppose instead that the variable was selected after a direct protected-class measure was removed, with communications describing it as a way to reach the same result. That evidence would support an inquiry into intentional proxy use. A third possibility is that the written model is neutral but loan officers override it selectively. The relevant evidence would then concern application, not merely model design. These are not automatic verdicts; they are different causal questions generated by the same outcome pattern.
The diagram's branches are not three mutually exclusive descriptions of every real case. An outcome disparity can coexist with evidence of intent, and a single lending process can contain both neutral components and explicit treatment. The point is to keep the causal questions separate. A result is an outcome; it is not automatically a motive. A correlation is evidence to interpret; it is not automatically an intentional classification.
This separation also explains why disparate treatment can reach a policy that looks neutral on paper. The FTC policy statement says investigators may look “behind the curtain” to determine whether a supposedly neutral explanation masks discriminatory intent. The CFPB's commentary makes the same practical point through its language about proxies. The law does not become satisfied merely because a lender avoids naming the protected group in the written rule.
What evidence can still do after the effects test
The effects-test change alters the role of evidence, not the existence of evidence. A disparity can trigger a question about a credit system. It can help identify which variable, business rule, human override, or communication deserves closer examination. It can also be part of a larger proof of intent. What the revised federal position rejects is the shorter inference that a neutral practice plus an unequal group result is enough, without more, to establish an ECOA violation on a disparate-impact theory.
That difference matters for model governance. A lender may compare approval rates, pricing, error rates, or other outcomes across groups for several reasons. The comparison might be used to find a profitable market, monitor risk, investigate intentional bias, comply with another law, or test whether a neutral policy has unequal effects. Those purposes are not legally interchangeable. A comparison used as an investigative lead asks what caused the pattern. An effects-test claim treats the pattern itself as the basis for liability. The new Regulation B text makes that distinction central under ECOA.
The rule does not create a complete catalog of permitted variables. It says a creditor may generally consider information so long as it is not used to discriminate, and it retains existing limitations on what information may be obtained or used. Its commentary continues to address empirically derived, demonstrably and statistically sound scoring systems and permits age as a predictive factor under specified conditions. That narrow example should not be read as a blanket approval for every protected-basis variable or every model that produces a statistically significant gap.
Nor does the rule answer every dispute about statistical proof. The CFPB declined to announce that a particular imbalance, peer comparison, or statistical technique is always sufficient or insufficient to prove disparate treatment. That restraint is important. Statistics may be weak because the sample is small, the comparison group is wrong, or the variables do not measure the decision at issue. They may be powerful when combined with evidence that a lender selected a variable for a prohibited reason or applied it selectively. The question is not whether numbers are admissible in the abstract, but what proposition they actually support.
The FTC policy statement illustrates the same caution from another angle. It criticizes Bayesian Improved Surname Geocoding, or BISG, and similar methods because they infer race from names and addresses when direct racial data are unavailable. The Commission argues that such methods add an inferential step and may have high error rates for racial minorities. That is the FTC's policy assessment of those methods, not a finding that all statistical analysis is invalid. A lender or investigator still has to identify the method, its assumptions, its error, its denominator, and the conclusion it can support.
Automated underwriting makes the distinction harder to see because a final decision may emerge from a chain rather than a single person's stated reason. Data are selected, transformed, weighted, validated, monitored, and sometimes overridden by staff. If a model produces unequal outcomes, the disparity may reflect a neutral risk relationship, a data-quality problem, a human policy choice, or an intentional design decision. The existence of automation does not eliminate intent, but neither does the existence of a disparity identify where intent lies. Evidence must connect the protected characteristic to a decision or design purpose.
Individual discretion creates a different evidentiary path. Two applicants may face the same formal score but receive different treatment through an exception, a manual adjustment, or a different explanation. A neutral policy can also be applied inconsistently to different groups. In those situations, records about the application of the rule may matter more than the rule's neutral wording. The CFPB's commentary preserves this possibility by focusing on criteria that are designed or applied as proxies, while leaving the ultimate sufficiency of proof to a fact-specific inquiry.
Good analysis also depends on what is being compared. An approval rate among people who completed a full application answers a different question from an approval rate among everyone who clicked an advertisement or began an inquiry. A pricing comparison may be distorted if applicants received different products, loan amounts, or risk profiles. A small sample may make a large percentage difference unstable, while a large disparity may still be uninformative if the comparator does not match the decision being studied. The final rule does not supply a universal statistical recipe, so a responsible conclusion has to name the population, the comparison, the time period, and the limits of the data.
Those details matter especially when a model is treated as a sealed object. A model's output is the end of a chain: a lender selects data, transforms variables, assigns weights, validates the system, monitors performance, and sometimes gives staff authority to override the result. Evidence at each point answers a different question. Validation may show how accurately the system predicts the outcome it was designed to predict; it does not, by itself, show why a lender chose that objective or whether a variable was intended as a protected-basis substitute. A record of human overrides may reveal a pattern that the published model cannot. Conversely, an unequal output may reflect the population or data used to train the system rather than a deliberate decision to disadvantage a group. The legal inference must not outrun the evidence at the stage where it was collected.
New evidence could therefore change the classification of the same disparity. A model-development document, a communication about the desired group result, a history of changes after a protected-class measure was removed, or a pattern of selective exceptions could strengthen an inference of intent. Better sampling, a more appropriate comparator, or evidence that the variable predicts a legitimate outcome across the relevant population could support a different explanation. None of those items automatically decides a claim, but each bears on whether the pattern is an effect, an intentional proxy, or an artifact of an inadequate comparison.
This is also why a lender should not treat the new rule as an instruction to stop examining outcomes. The agencies' own account says the legal landscape remains layered. Other federal and state laws may use different standards, and a pattern that is not a standalone ECOA effects-test violation may still be relevant to another claim or to an investigation of intentional treatment. The governing question must therefore identify the statute, the transaction, the protected basis, the decision process, and the purpose of the evidence before drawing a conclusion.
A careful reader can now make a limited but useful judgment. If a facially neutral policy produces a disparity and nothing more is shown, the revised federal ECOA interpretation does not treat the disparity alone as the effects-test violation that the former Regulation B language described. If evidence shows that the policy was chosen or applied to disadvantage a protected group, the fact that the policy looks neutral does not end the analysis. If the evidence only shows correlation, the next step is inquiry, not automatic condemnation. Whether the connection is sufficient remains a question for the facts and, where disputed, a court.
That unresolved line leads to the next question: why did the CFPB decide that ECOA's text supports this intent-centered distinction when earlier Regulation B language, legislative history, and some judicial readings supported an effects test? Understanding that dispute requires moving from the mechanics of a credit decision to the sources of statutory meaning.