The Algorithm Remembers the Red Line
In the 1930s, the government drew maps to keep Black families from getting mortgages. The maps are gone. The data isn't. And AI is drawing new ones.
In July 2025, the Massachusetts Attorney General fined a student-loan company called Earnest Operations $2.5 million for discriminating against borrowers from historically Black colleges.
Earnest didn’t use race in its algorithm. It didn’t have to.
The company’s AI underwriting model included a variable called the Cohort Default Rate: the average student-loan default rate for the applicant’s school. HBCUs and schools serving low-income Black and Hispanic students have higher CDRs. Not because their graduates are worse borrowers. Because the schools serve populations that carry more student debt, earn less in entry-level jobs, and have fewer family resources to absorb early-career financial shocks.
The algorithm didn’t see race. It saw a number that was shaped by race. Every borrower from those schools was penalized regardless of their individual creditworthiness. The model also automatically denied anyone without at least a green card, cutting out DACA recipients and visa holders before the risk model even ran.
The Massachusetts AG called it what it was: unfair treatment of historically marginalized borrowers. But Earnest’s model wasn’t unusual. It was standard practice that happened to get caught.
The Maps Are Gone. The Math Isn’t.
In the 1930s, the federal government’s Home Owners’ Loan Corporation drew maps of American cities. They color-coded neighborhoods by perceived lending risk. Green for “best.” Blue for “still desirable.” Yellow for “declining.” Red for “hazardous.”
The red neighborhoods were almost always Black. The maps became policy. The Federal Housing Administration refused to insure mortgages in redlined areas. Banks followed. For three decades, the color of the lines on a federal map determined who could build wealth through homeownership and who couldn’t.
The Fair Housing Act of 1968 made explicit redlining illegal. The Equal Credit Opportunity Act of 1974 banned discrimination in lending. The Community Reinvestment Act of 1977 required banks to serve the communities where they took deposits.
The laws worked on the surface. You can’t draw a red line on a map anymore.
But the data remembers.
A 2026 study in Crime & Delinquency, using data from 48 police departments, found significant increases in present-day violent crime, property crime, and weapon offenses in neighborhoods that were redlined in the 1930s, even after controlling for historical differences. Research from SAVI found that HOLC grades alone explain 62% of the variation in violent crime rates between census tracts and 53% of household income variation. Redlining shaped poverty. Poverty shaped crime. Crime shaped policing. And now policing data shapes the algorithms.
The Loop That Launders the Bias
Here’s what I keep coming back to.
Think of it as a washing machine for discrimination. Each cycle makes the original stain harder to see.
Cycle one: cameras are placed where police say crime is highest. In Hampton Roads, Virginia, that means majority-Black census tracts get four times the cameras. The CNU study I wrote about in my last piece documented this. Nobody called it racial targeting. They called it data-driven policing.
Cycle two: more cameras generate more data. More data means more alerts, more stops, more recorded incidents. The neighborhood’s “crime score” rises. Not because crime increased. Because measurement increased.
Cycle three: that score leaves policing. Property valuation algorithms, the models that estimate what your home is worth, incorporate crime statistics, school ratings, and neighborhood-quality indicators. Zillow’s Zestimate draws from these variables for more than 200 million monthly visitors. HUD research confirms that these automated valuation models produce systematically larger errors in majority-Black neighborhoods. The training data, built from historical sale prices, embeds the decades of undervaluation caused by redlining.
Cycle four: insurance follows. The Consumer Federation of America found that homeowners with lower credit scores pay roughly $1,996 more per year in premiums. Twenty-two percent of Native American homeowners are uninsured. Fourteen percent of Hispanic homeowners. Eleven percent of Black homeowners. Six percent of white homeowners. ProPublica documented auto insurers charging up to 30% more in minority ZIP codes with similar accident risk.
Cycle five: tenant screening closes the loop. SafeRent Solutions settled a $2.275 million class action after its algorithm disproportionately failed Black and Hispanic housing-voucher holders. The model relied heavily on credit history and ignored that vouchers cover most of the rent. Between 57% and 90% of landlords now use algorithmic screening reports. RealPage markets an “AI Screening” product that predicts “willingness to pay” using 30 million lease records.
By the time a mortgage underwriter, an insurance adjuster, or a landlord sees your neighborhood’s score, the original source of the bias is invisible. It looks like math. It looks like risk assessment. The algorithm never saw race. It didn’t need to.
Advocates call this data laundering. Using proxies like ZIP code, credit score, and prior addresses to reproduce racial outcomes without ever naming race, then claiming neutrality.
The Guardrails Are Coming Down at Exactly the Wrong Moment
For decades, the primary legal tool against algorithmic discrimination was disparate impact: the principle that a policy can be discriminatory even without discriminatory intent, if it produces discriminatory outcomes. You don’t have to prove the algorithm is racist. You have to prove the results are.
On April 23, 2025, President Trump signed Executive Order 14281, directing all federal agencies to “deprioritize enforcement of all statutes and regulations to the extent they include disparate-impact liability.” In December 2025, the DOJ rescinded Title VI disparate-impact regulations entirely. HUD proposed withdrawing Fair Housing Act disparate-impact regulations in January 2026.
The CFPB closed all fair-lending investigations based on statistical analysis. The EEOC directed staff to close pending disparate-impact cases. The Consumer Financial Protection Bureau finalized a rule eliminating disparate-impact analysis under the Equal Credit Opportunity Act.
The only federal AI-specific law enacted so far is the TAKE IT DOWN Act, which criminalizes nonconsensual deepfakes. Nothing on credit. Nothing on housing. Nothing on insurance. Nothing on neighborhood scoring.
The Colorado AI Act, the most ambitious state effort, has been delayed twice and may be rewritten before its AG even begins enforcement. The EU’s AI Act classifies credit scoring as high-risk, with requirements taking effect in August 2026. But that covers European markets, not American ones.
State attorneys general are stepping into the gap. Massachusetts, Oregon, New Jersey, New York, and Connecticut have all signaled they’ll use existing consumer-protection and fair-housing laws against AI bias. But state enforcement is patchwork. If your state AG isn’t filing suits, the algorithm runs unchecked.
What I See When I Look at This
I’m an IT Director. I think about data flows the way a plumber thinks about pipes. Where does the data come from? Where does it go? Who can see it? Who controls it?
When I look at neighborhood scoring, I see a plumbing problem. Surveillance data flows into crime databases. Crime databases flow into neighborhood-quality scores. Neighborhood-quality scores flow into insurance pricing, property valuations, and lending models. Each system treats the incoming data as objective. None of them audit the source.
The DOJ settlement with RealPage required the company to restrict its training data to records at least 12 months old and to stop making geographic predictions narrower than statewide using cross-owner data. That’s a start. It’s an acknowledgment that the data itself is the problem, not just the decision.
But RealPage is one company. The same data infrastructure feeds hundreds of vendors. And none of them are required to trace where their training data came from or test whether it reproduces historical discrimination.
What I Don’t Have Answers To
I don’t know how to build a lending model that accounts for neighborhood context without reproducing neighborhood bias. Removing ZIP code doesn’t work. The bias routes through a dozen other variables that correlate with the same geography. Credit score. School district. Property condition. Prior addresses. Each one carries the signal that ZIP code carried.
I don’t know whether the EU’s approach, classifying credit scoring as high-risk and requiring bias testing, will actually produce better outcomes or just produce better paperwork. The compliance deadline is August 2026. We’ll see.
And I’m not sure state AGs can scale fast enough. The Massachusetts Earnest settlement was landmark. But Earnest is one company, and $2.5 million is a rounding error in the lending industry. The algorithmic tenant screening market alone covers 60 to 90% of rental applications. You can’t enforce your way out of a problem that operates at that scale.
Same Map, New Math
In the 1930s, a government bureaucrat drew a red line around a neighborhood and wrote “hazardous” in the margin. That line determined who built wealth and who didn’t for three generations.
The maps are archived now. You can view them online. They look like relics of a different era.
But the data they created is still flowing. Through crime statistics trained on decades of over-policing. Through property valuations trained on decades of undervaluation. Through credit models that penalize the schools your parents could afford and the neighborhoods your grandparents were allowed to live in.
The algorithm doesn’t know about the red line. It doesn’t need to. The red line is in the training data.
If you’re thinking about these questions too, I hope you’ll subscribe.
Rachel Ankerholz is an IT Director and writer exploring the intersection of AI ethics, accessibility, and human-centered technology. She writes about who gets included, and who gets left behind, when we build systems.