On July 21 a rule the CFPB finalized in April takes effect and removes the “effects test” from Regulation B, the rule that implements the Equal Credit Opportunity Act. The agency did not narrow the test. It deleted it, and added a flat statement to the rule text: the Act “does not provide that the ‘effects test’ applies for determining whether there is discrimination in violation of the Act.”
That single edit pulls the primary federal mechanism for catching lending discrimination that no one intended. It lands as algorithmic underwriting becomes the default.
What the effects test actually did
Fair-lending law recognizes two kinds of discrimination. Disparate treatment is intent: a lender that prices or denies differently because of a protected class. Disparate impact, the effects test, is outcome: a facially neutral policy that falls harder on a protected class and cannot be justified by business necessity. No intent required.
The second is the one that bites model-driven lending. A machine-learning underwriter is trained to optimize approval and default prediction against historical data. It has no intent. It finds correlations. Some of those correlations are proxies for race, sex, age, or national origin, encoded in ZIP code, education level, spending pattern, or device type. The model does not know it is discriminating, and often neither does the lender. Disparate impact was the doctrine built precisely to surface that: measure the outcome, ignore the intent.
Strip it from the federal statute and you strip the federal lever for auditing a model by its results. What remains at the CFPB is disparate treatment, which requires evidence that someone meant to discriminate. Against a gradient-boosted model with thousands of features and no author of intent, that evidence is close to unobtainable. Operationally, the change is quiet but large: a lender no longer has to justify a disparate approval rate to the Bureau, no longer has to prove business necessity for a variable that correlates with a protected class, and no longer has to run the outcome test that used to precede a model going live. The federal question shifts from “does the model produce a disparate result” to “can anyone prove you wanted one.”
The CFPB’s reasoning
The agency’s case is textual, not empirical. It argues the prior reading, that disparate-impact claims are cognizable under ECOA, “is not the best interpretation” of the statute, because it leaned on legislative history rather than the words Congress enacted. The Bureau told Congress this spring it had already stopped using disparate impact in supervision and would instead look for direct evidence of intent. The rule codifies a practice already in place. The same document also narrowed the “discouragement” standard and tightened conditions on special purpose credit programs, the two other fair-lending softenings that shipped alongside.
Why “off the hook” is wrong
Here is where the client alerts stop and the real exposure map starts. The effects test is gone from ECOA. It is not gone from lending.
Mortgage and housing credit answer to the Fair Housing Act, administered by HUD, and the FHA’s disparate-impact standard survives this rule untouched. HUD’s disparate-impact regulation at 24 C.F.R. 100.500 is still in force, and the Supreme Court upheld FHA disparate-impact liability in Texas Department of Housing v. Inclusive Communities, subject to a strict causality test. An AI mortgage model that produces disparate approvals still carries outcome-based liability. It just runs through HUD and the courts now, not the CFPB. HUD has proposed rescinding its rule, but that comment period does not close until 2027, so the exposure is live today.
State law is the second surviving axis, and it is widening, not shrinking. New York and Illinois have fair-lending statutes read to permit disparate-impact claims independent of Regulation B. New Jersey codified disparate impact and issued guidance aimed directly at algorithmic decisioning, with explicit demands for AI explainability and governance documentation that exceed anything the federal rule now requires. Enforcement is not theoretical: a state attorney general has already settled a fair-lending action built on the outputs of an AI underwriting model, using outcome analysis to do it.
What actually changed for a lender
The practical shift is jurisdictional, not a release. A national lender that ran one disparate-impact analysis to satisfy the CFPB now faces a split map. For non-mortgage consumer credit, federal outcome-based liability is gone, but a growing patchwork of state statutes fills the gap unevenly, and the strictest states demand more model documentation than the federal rule ever did. For mortgage credit, nothing has changed: FHA disparate impact still governs the model, through HUD and through Inclusive Communities.
The compliance function that treated fair-lending testing as a single federal checkbox is the one most exposed. The rational response is not to stop outcome testing. It is to keep testing, because HUD and a half-dozen state regulators still read the results, and because a model you have stopped auditing for disparate outcomes is a liability you can no longer see. The federal alarm went quiet on July 21. The fire code did not change.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.