Imagine a physical loan officer’s desk in 1962, with a map of the city tacked to the wall behind it, some neighborhoods marked in red marker, and the implicit knowledge that applications from those areas would not be accepted, regardless of the applicant’s financial situation. Redlining was eventually outlawed. It is regarded as one of the most significant and intentional instances of housing discrimination in American history, a methodical way to keep Black families from accumulating wealth for future generations through homeownership.
It was intended to be abolished by the Fair Housing Act and the Civil Rights Act of 1968. The desk has vanished. The map has vanished. The algorithm has taken the place of the red lines, making it more difficult to see and far more difficult to challenge.
Nowadays, most mortgage lending decisions in the US are made or significantly influenced by automated underwriting algorithms. They receive inputs across hundreds of financial factors, process applications more quickly than any human reviewer, and produce recommendations—approve, refuse, or flag for manual review—with apparent accuracy and speed that gives the outcome an air of objectivity. The machine made the decision. However, past loan data was used to train the system.
Additionally, decades of discriminatory behaviors are shown in historical lending statistics. With statistical certainty, a system trained to identify “creditworthy” patterns by examining previous approvals will learn to mimic the traits of those who were previously approved, who were disproportionately and intentionally white applicants. It is not necessary to expressly code the prejudice. It shows up in the training set.
One of the most thorough analyses of this issue in recent years was carried out by The Markup and the Associated Press, which examined what happened to applications from various demographic groups after controlling for financial characteristics, such as debt-to-income ratios, loan-to-value ratios, and the variables that lenders claim are the only ones that matter. The outcomes were noteworthy enough to warrant a clear statement.
Algorithms were 80% more likely to reject Black candidates, 70% more likely to reject Native Americans, 50% more likely to reject Asian and Pacific Islander applications, and 40% more likely to reject Latino applicants, even when all other factors were held equal. The result was determined by anything other than the finances. Without being requested to search, the algorithm was locating it.
The “black box” issue contributes to the difficulty of solving this. One of the most tightly protected types of intellectual property in the lending sector is proprietary underwriting software. The precise weights given to particular factors, the way inputs interact, and the thresholds that result in rejection are all not publicly known, and existing legal frameworks do not mandate that they be.
Even when the statistical pattern is clear, it is nearly impossible for federal regulators looking into a possible Fair Housing Act violation to prove discriminatory intent in a legal sense because they can see the results but not the mechanism causing them. The prejudice is apparent. It is not the proof needed to take action on it.
Everything is compounded by the credit score system. Medical debt is a financial burden that disproportionately affects lower-income households. Older FICO models, which are still mandated by Fannie Mae and Freddie Mac in the majority of lending scenarios, penalize medical debt while providing credit for forms of borrowing that have historically been more accessible to white Americans. They completely disregard rent payments.
A mortgage lender does not give credit to someone who has paid their rent on time each month for ten years—exactly the kind of financial dependability they say they are searching for. Someone who has had credit card debt since college does. It’s difficult to ignore how frequently the scoring system disadvantages the same groups that have historically been affected by discriminatory lending practices. At the very least, it’s worth asking loudly and frequently whether that’s a coincidence or an artifact.
