Name-Blind Screening Under-Delivers, and the Field Trials Say Why

Does removing the name reduce bias?

Not reliably, and the randomized trial of the practice points the other way. When the French public employment service randomly assigned about 600 participating firms to receive anonymized or name-bearing applications, the interview gap widened: minority candidates' interview rate fell and majority candidates' rose ¹. The firms had volunteered, and volunteers were already interviewing relatively more minority candidates than others.

The authors name two mechanisms and both are about that setting. Firms selected into the program, so the sample was drawn from employers already inclined to interview these candidates. And stripping the name also strips the context a recruiter was using to discount a negative signal, such as an employment gap, for a candidate they thought they could place. Remove the context and the discount goes with it.

The result belongs to its setting. It is a program evaluation in France, on firms that volunteered, and it did not test automated screening or establish a general law. What it establishes is narrower and still useful: removing an identifier is an intervention with its own selection effects, not a neutral safety measure that can only help.

The evidence usually cited on the other side does not say what the pitch says it does. Bertrand and Mullainathan sent 4,870 fictitious resumes to help-wanted ads in Boston and Chicago and found resumes with White-sounding names drew a callback 9.65 percent of the time against 6.45 percent for otherwise equivalent resumes with African-American-sounding names, a gap of 3.20 percentage points, or 50 percent ². It is a real and careful result. It is also a callback measurement from 2001 and 2002 in two cities, and the authors write plainly that they cannot translate it into gaps in hiring rates or earnings ².

What survives redaction?

School, postcode, dates, employment gaps, visa phrasing, the names of former employers, and the register of the writing itself. Each one correlates with something the redaction was meant to hide, and each stays on the page. A correspondence study varies the name and holds everything else identical, which is exactly what a real application does not do.

The employment gap is the clearest case, because it is priced at the screen and almost nobody thinks of it as an identity signal. A correspondence study sending roughly 12,000 resumes to real online postings across the 100 largest US metropolitan areas found the callback rate falling from roughly 7 percent at one month of unemployment to roughly 4 percent at eight months, about 45 percent lower, after which additional months made almost no further difference ⁴. Fieldwork ran in 2011, and the authors read the spell as a signal employers use rather than as skills decaying.

Thinning a document is not neutral either, once a model is doing the reading. In a study of resume screening through language-model retrieval, cutting resumes down to a name and job title produced significant race differences in more of the bias tests than full-length resumes did, and the group selection-rate gaps widened ⁵. The paper's own percentages in that section do not reconcile to whole numbers of tests, so read the direction and leave the decimals alone.

Those documents kept the name and lost the content, which is the reverse of what an anonymization tool does, so what the study measures is thinning. The mechanism is the part that carries over: strip a document down and whatever identity signal is left takes up more of what the model has to read. A thin profile fed to a ranking layer is not automatically the safer object, and school, postcode and dates are usually still on it. That two-layer pipeline is worth understanding before configuring anything: what AI resume screening does to the pile before you open it.

Which stage does your gap open at?

Only your own pass rates can tell you, and the field-experiment average says the screen is not the whole of it. Across the twelve field experiments that follow applicants past the interview invitation, majority applicants received 53 percent more callbacks than comparable minority applicants and 145 percent more job offers ³. Roughly half the discrimination in offers opens after the resume screen, in rounds no redaction reaches. Blinding acts on one stage only.

Those numbers need their limits attached. Twelve studies is a small and selected pool, skewed toward roles that can be audited in person or by telephone, and 145 percent is a ratio of offer rates rather than a percentage-point difference. The additional discrimination after the interview correlates only weakly with the callback gap, so a firm cannot read its own later-stage behavior off its screening numbers in either direction.

Which is the operational point. Before buying anonymization, compute pass rates stage by stage for the groups you can lawfully count, and see where your own funnel narrows. What you may collect, and how the analysis is held once it exists, turns on the jurisdiction the role sits in, so settle that with counsel before the first query runs. If the screen passes people evenly and the onsite does not, blinding the screen changes nothing you care about, and the money and the political capital have been spent on the wrong stage.

Two adjacent pieces do this arithmetic properly: finding which stage the gap opens at and measuring bias stage by stage. Both are more work than a toggle and both tell you something a toggle cannot.

Replace blinding with criteria you wrote down first

Write the criteria before the pool arrives, apply them one at a time, and record a reason for each rejection in the words of the criterion. That is slower than a toggle, and it acts on the part blinding cannot reach, which is what a reviewer is free to weigh once the name is gone. Criteria written after reading fifty applications describe those fifty people.

Then measure where your own hiring happens rather than assuming a background rate. In an audit sending more than 83,000 applications to entry-level vacancies at 108 of the largest US employers, contact gaps were highly concentrated: the top quintile of discriminating firms accounted for nearly half the contacts lost to Black applicants, with a Gini coefficient of about 0.4 ⁶. Concentration is not a clean bill of health for everyone else, and the authors say they could not reject the hypothesis that all 108 firms weakly favored White names. What it does establish is that this is a property of particular employers and particular jobs, which means it has to be measured where you hire.

Three habits carry most of the value, and none of them needs a vendor. Decide what a rejection reason may say before the first application is read. Keep the reasons in a field somebody can count later. And review the criteria themselves for the ones that are doing work you never intended, which is the usual home of a gap nobody chose.

Where the screen has stopped separating people at all, the problem is upstream of fairness and is its own question: what to screen on when every resume looks perfect, and how to tell whether the screen is throwing away the wrong people.

See a sample report >