“Covert bias” and American homeowners

It’s not an entirely new discussion — the concern of AI and its embedded racist tendencies. It’s not even so much a discussion, as opposed to a known truth; AI is built by people, people hold racial prejudices. It’s as simple as that.

But the discussion around how we combat that truth — that’s the thing I’m most interested in.

Even though it’s been more than a year now since my core user-base was homeowners (in my time as Senior Content Designer with Rocket Mortgage), I still find myself wondering about the things impacting them today.

Lately, I’ve found myself wondering something in particular:

If AI is inherently racist/at risk of perpetuating racism in its outputs, how is that impacting homebuyers and homeowners of color in America today?

So, I began researching. I reviewed recent studies, devoured articles, and found myself scribbling frantic notes in my notepad. The answers I found were not happy ones.

I was quickly reacquainted with the reality of generative AI tools and their impact on people of color through “covert racism,” with one source speaking directly to the existence of prejudice based on dialect.

From the way we speak…

It noted that dialect prejudice is fundamentally unique to other racial biases that have been studied in language models — specifically because when interacting with LLMs, the speakers’ race is likely never made clear. This ties back to previous research on auditory profiling.

“For example, researchers have previously found that landlords engage in housing discrimination based solely on the auditory profiles of speakers, with voices that sounded Black or Chicano being less likely to secure housing appointments in predominantly white locales than in mostly Black or Mexican American areas.”

Furthermore, it stated the exact notion I mentioned earlier:

“The fact that humans hold these stereotypes indicates that they are encoded in the training data and picked up by language models.”

To the way we’re assessed…

Another article spoke more to the impact of AI on American homebuyers, citing a study from the University of California, Berkeley.

The study showed that “an AI mortgage system routinely charged black and lantinx borrowers higher rates for the same loans compared to their white counterparts.”

“The current housing crisis disproportionately affects Black, Latinx, Indigenous, and low-income households: the same groups that are most affected by biased algorithmic decision-making tools.”

The source ends with an astute perspective: the very tools made with the promise of modernizing the “slow and inefficient homeownership” process are the same things making housing discrimination easier than ever before.

There are verified roadblocks

The last article focused on the racial bias in mortgage underwriting decisions, citing a study done by Lehigh University (2024).

“In an experiment using leading commercial large language models (LLMs) to evaluate loan application data, Lehigh researchers found that LLMs consistently recommended denying more loans and charging higher interest rates to Black applicants compared to otherwise identical white applicants.”

“This finding suggests that LLMs are learning from the data they are trained on, which includes a history of racial disparities in mortgage lending, and potentially incorporating triggers for racial bias from other contexts,” said Donald Bowen III, one of the authors of the study.

The results of the study were clear: Black applicants consistently faced higher barriers to homeownership, even when they were financially the same as their white counterparts. Models also showed bias against hispanic applicants.

That’s not an abstract statistic. That’s a family paying more for the same house, for no reason but the one they can’t change. Not a glitch. The system working as trained.

So, what do we do about it?

Here’s where things get complicated. When I started digging into how regulators were responding to biased AI in lending, I expected to find a growing wall of protection. Instead, I found a landscape in flux.

For years, the Consumer Financial Protection Bureau (CFPB) held a clear line: new technology doesn’t exempt lenders from federal consumer protection laws, and courts have ruled that choosing to use a biased algorithm can itself create legal liability. That was a real foothold for anyone harmed by a discriminatory model.

That foothold has been shifting. The CFPB’s operations were shut down in February 2025 amid a federal pivot toward deregulation, and by May 2026 the agency had finalized a rule narrowing its fair lending enforcement — a retreat from its more expansive prior approach. Oversight hasn’t disappeared entirely (that same month, the CFPB reminded lenders they still owe applicants specific reasons for AI-driven denials), but the federal direction is toward less scrutiny, not more.

Where the federal government is stepping back, some states are stepping in — Colorado, Michigan, New York, and California have all introduced their own AI-lending rules. Globally, the EU’s AI Act will require bias testing and human oversight for credit and loan decisions starting August 2026.

So what does that mean for the rest of us — the homebuyers, the advocates, the people building these tools?

It means the fight against algorithmic housing discrimination can’t rely solely on regulators to carry it. It has to happen on multiple fronts at once:

Inside the industry: Lenders and tech companies need to run their own bias audits before models ever touch a real application — not after a lawsuit forces the question.Inside the data: If a model is trained on decades of discriminatory lending history, it will learn discrimination as a pattern to repeat, not a mistake to avoid. Actively curating and correcting training data is a starting point, not a nice-to-have.Inside the room: Diverse product, design, and engineering teams — the people actually building these tools — remain one of the simplest, least regulated defenses. It’s a lot harder to ship a biased assumption when someone in the room recognizes it as one.Outside the industry: Journalism, research, and public pressure — like the studies I’ve cited here — are often what forces a reckoning that regulation hasn’t caught up to yet.

None of this fixes the problem overnight. But it does mean the responsibility isn’t just sitting with a government agency that may or may not have the appetite to enforce it.

It’s sitting with the people who build these systems, the people who study them, and the people — like me — who used to help homebuyers navigate a process that was already hard enough before an algorithm got a vote.