New Fair-Lending Rule Leaves AI Mortgage Lenders Exposed This July

What the April rule actually changed

The document is narrower than the early write-ups suggest. The Bureau finalized the rule under Acting Director Russell Vought, and the official citation runs 91 FR 21620. It drew roughly 64,500 comments. Three things shifted: disparate impact came out, the rules on discouraging would-be borrowers were tightened, and special-purpose credit programs, the vehicles lenders use to steer credit toward underserved groups, now face new limits.

Disparate impact is the idea that a lender can violate fair-lending law without meaning to, because a neutral-looking policy falls harder on one protected group than another. For about 50 years that theory carried real weight under ECOA. The new rule removes it. The CFPB now states that the law does not authorize disparate-impact liability, and it deleted every reference to the so-called effects test. Statistical disparity, on its own, no longer proves an ECOA violation.

One caveat survives, and it matters. A policy that looks neutral but works as a stand-in for race or sex, and is applied to harm people on that basis, still counts as intentional discrimination. Proxy discrimination remains illegal. Accidental disparity does not.

Reaction split along the lines you would expect. Lawyers at one firm read the rule as a disciplined effort to bring enforcement back in line with what the statute literally says. Jesse Van Tol, who runs the National Community Reinvestment Coalition, told American Banker the change was “a major step back.”

The rule did not arrive in a vacuum. It carries out an April 2025 executive order directing federal agencies to scrap disparate-impact liability wherever the law allowed. The CFPB was one of several agencies told to move in the same direction, which is part of why lenders should not read the change as the Bureau’s own quiet reassessment of fair-lending risk. It is policy flowing downhill from the White House, and policy can reverse with an administration. The statutes underneath it, ECOA aside, do not.

Why the liability did not actually leave

The CFPB administers one fair-lending statute, ECOA. It does not administer the others, and that is the point lenders need to hold onto.

Disparate impact also lives under the Fair Housing Act, a separate law the Bureau cannot rewrite by amending Regulation B. The Supreme Court upheld disparate-impact claims under the Fair Housing Act in 2015, in Texas Department of Housing and Community Affairs v. Inclusive Communities Project, and that ruling stands. A mortgage decision that pushes a protected group toward worse terms can still draw a Fair Housing Act claim, whatever Regulation B now says about intent.

State law adds another layer. During the comment period, a coalition of state attorneys general, including those of California, New York, Illinois and Massachusetts, filed a joint letter opposing the change. They urged the Bureau to keep the existing disparate-impact provisions and warned the rewrite would lead to more discrimination, not less. Many of those states run their own fair-lending statutes, which the CFPB rule does nothing to soften.

A lender that treats the April rule as a green light has swapped one regulator for several. The federal agency on the ECOA beat has pulled back. The Fair Housing Act, the state attorneys general and the plaintiffs’ bar have not.

The firm that tracks this closely flagged that the rule sets up a collision with decades of agency practice, several court decisions and other fair-lending regimes, and said it expects legal challenges. Even people who welcome the rule expect it to be fought over.

Why AI raises the stakes

The exposure is sharpest for lenders that rely on machine learning, because of how those systems behave over time.

A traditional credit model is largely static. You build it, test it, and review it periodically. A machine-learning system retrains on new data and shifts as it goes. A model that looked clean on the day it launched can, months later, start leaning on a variable that tracks race or national origin, without anyone choosing that outcome. The standard safeguard is unglamorous: running the model’s results through fair-lending tests, checking whether approvals or pricing skew along protected lines, and documenting what turns up. Under the old framework, that drift was a disparate-impact problem lenders had to monitor. A lender that stops monitoring now, on the theory that disparate impact is dead, has misread the risk. The Fair Housing Act exposure from the same drift remains.

Fannie Mae and Freddie Mac moved in the opposite direction from the CFPB. In April, Fannie Mae issued Lender Letter LL-2026-04, a governance framework for any seller or servicer using AI or machine learning to originate or service its loans. It takes effect August 6. The letter runs two pages and is strict. Lenders must keep written policies covering the full life of every AI system they use, review them at least once a year, and hold their vendors to the same standard. On request, they must tell Fannie Mae what the tools do and what safeguards are in place. Freddie Mac got there earlier, with rules in force since March.

Put the two together and the timeline is telling. The federal consumer regulator narrowed its rules effective July. The companies that buy most of America’s mortgages tightened theirs effective August. The Fannie letter’s message is plain: if you use AI, you govern it, document it, own it and prove it on demand. That is not an industry that believes fair-lending risk has gone away.

Warning: Dropping fair-lending model testing because the CFPB rolled back disparate impact under Regulation B leaves Fair Housing Act and state-law exposure fully intact, and nonbank lenders have the least cover of all.

Nonbanks carry the most exposure

Some lenders are more exposed than others. Banks answer to prudential supervisors, the examiners who inspect the machinery regardless of what the CFPB is doing in a given quarter. Nonbank mortgage lenders, which now originate a large share of U.S. home loans, do not have that backstop. They felt the CFPB’s presence more directly because fewer other regulators were watching.

With the CFPB pulling back, the gap widens. A nonbank that treats the April rule as license to thin out its fair-lending testing has the least cover if a Fair Housing Act claim or a state action arrives. It still has to satisfy Fannie and Freddie, since selling loans to them is the business model. The relief is loudest exactly where the remaining obligations bite hardest.

The history here is worth keeping in view. Disparate-impact theory has been part of fair-lending enforcement for roughly half a century, and a large share of the redlining and pricing cases brought against lenders over that period leaned on it. Borrowers who believe an algorithm priced them unfairly have not lost the ability to sue; they have lost one federal forum for it. The Fair Housing Act route runs through the courts and the Department of Housing and Urban Development, neither of which answers to the CFPB. For a lender, the practical question is not whether the theory still exists. It is which door a claim now comes through.

There is a second point for the nonbank market. A broker who never deals with Fannie directly can still be pulled in, because the Fannie letter requires sellers and servicers to hold their vendors, subcontractors and third-party originators to the same governance standard. Wholesale lenders are likely to write AI disclosure terms into their broker agreements. The discipline the CFPB lifted at the front door re-enters through the supply chain.

The supervisory gap predates this rule

The confusion is not new. In December, the Government Accountability Office warned that as proptech and AI spread through homebuying, the Federal Housing Finance Agency had left Fannie and Freddie without clear written direction on how to comply with fair-lending law. The GAO asked the FHFA to fix it. The agency, per the report, neither agreed nor disagreed. So the question of who enforces what, against which AI system, predates the April rule and will outlast it.

The Bottom Line

Federal deregulation is not deregulation of risk. If you run AI underwriting, the practical steps are unglamorous: keep testing your models for biased outcomes, keep the documentation Fannie and Freddie will ask for, and treat the Fair Housing Act and your state attorney general as the audience that still matters. The CFPB eased off. The other enforcers did not.