In the News

THE ENDURING SYMBOL of Justice as a woman wearing a blindfold to signify impartiality was created long before anyone dreamed of eradicating discrimination based on skin color. Yet the statue well depicts the notion that eliminating human discretion can eliminate racial discrimination and, eventually, racial disparities.

Give judges rules for sentencing criminals, and they can't be harsher to blacks than whites. Let a computer decide whether to grant a mortgage, and its decisions can't be biased.

Vowing to reduce "unwarranted sentencing disparities," Congress created rigid sentencing guidelines for federal judges in 1984, later supplemented with mandatory minimum sentences. Today, black offenders get longer prison sentences than whites.

Eager to cut costs, big mortgage lenders are automating the process of applying for a mortgage. Fannie Mae, the financial giant that buys mortgages from banks and brokers that originate them, runs half of the six million applications it receives annually through a computer called Desktop Underwriter. It is colorblind, but rejects black applicants more often than white; the company won't give precise figures.

These very different cases illuminate an issue made increasingly important by what Christopher Edley, a professor at Harvard Law, calls "the inexorably grinding engines of technology and information." Does reducing the direct role that humans have in decision-making reduce racial discrimination or perpetuate racial disparities by hard-wiring them?

Because Federal sentencing rules aren't -- yet -- locked up in computer code, they are easy for all to see. In a frenzy of fright over crack cocaine and without racial animus, Congress assigned the same mandatory five-year prison sentence for trafficking in 500 grams of powder cocaine or five grams of crack cocaine.

The rub: 85% of those sentenced for crack are black and 68% of those sentenced for powder cocaine are not.

The lesson: Rules, whether embodied in law or software, are the product of human discretion. And rules that aren't motivated by racism can have an unintended, disparate impact on blacks.

THE SAME LESSONS apply to computer models that Fannie Mae and competitors use to predict almost instantly how likely a mortgage applicant is to default. Loan takers punch into a computer factors ranging from credit-card history to down-payment size -- but not race. Using past experience as a guide, the computers OK about 70% of applicants. The rest are kicked out for human review; most are eventually rejected.

Lenders argue that the computers allow them to make more loans. They know that many would-be borrowers whom they reject would make monthly payments but get lumped with likely deadbeats because they resemble each other. Computers see distinctions that humans don't -- and more often get to yes.

To make the point, another big lender, Freddie Mac, asked both human and computer underwriters to review 1,000 applications for loans made to low- and moderate-income homebuyers. The humans graded 51.7% as "accept"; Freddie Mac's computer model accepted 85.2%.

Yet computers still reject blacks more often than whites because black applicants tend to make smaller down payments, have more blemishes on their credit histories and have less money in the bank. Statistical analysis says these three traits lift the risk of default. Critics say they reflect the long-lived effects of racism and embed them in computer code.

One solution: Find clues to predict the odds of repayment that don't disadvantage blacks or other minorities. One leading possibility is a registry to track rent payments. The theory is that folks who pay rent on time are likely to make mortgage payments on time; blacks are more likely to be renters than whites.

BUT WHEN SHOULD HUMANS intervene to cull factors from the computer model that hurt blacks? People who borrow from finance companies are slightly more likely to default on their mortgages so this is a feature in some computer models. But blacks are much more likely to borrow from finance companies than whites. Concluding that the large cost (to black applicants) outweighs the benefits (fewer defaults), Fannie Mae says it no longer uses this factor.

And what about the cases the computer refers for human review, a pool in which black applicants are more likely to land than white? Upbeat lenders say using computers for easy calls allows humans to spend more time finding ways to approve more iffy cases.

Perhaps. But as automated underwriting spreads, the only loans that humans will evaluate are those the computer rejected. "These loans are starting to develop a stigma," cautions Regina Lowerie, chief executive of Gateway Funding Diversified Mortgage Services Inc., a Philadelphia mortgage company. That reopens all the issues of human bias that use of the computer removed initially.

Artists didn't blindfold Justice routinely until the 16th century. The addition may not have been meant to suggest impartiality, but rather to attack courts' tolerance of abuse of the law. At the Supreme Court's building, some likenesses of Justice have blindfolds; some don't.

The unintended lesson: We'll be relying on both race-blind rules and race-aware humans for a long time.

-- David Wessel