Algorithmic monoculture in hiring

Researchers at Stanford, Chapman and Northeastern obtained something nobody outside the industry had held before: the actual decisions of a deployed hiring algorithm at scale. The data covered 3.4 million real applicants submitting 4 million applications to 156 employers across 11 market sectors, every one of them screened by algorithms from a single vendor. That arrangement, where many employers rely on the same automated screener, is what they call an algorithmic monoculture. Over 60 percent of the Fortune 100 use one vendor's tools.

The question they asked was whether a shared screener produces shared rejections. It does. Among applicants who submitted four applications, 10 percent were rejected everywhere, significantly more than a model of independent decisions predicts. The comparison is what makes the finding hold. They tested the same independence baseline against a large earlier study in which 83,000 applications were sent to 108 large firms without a shared algorithm, and there the baseline predicted the real rejection rates accurately. So the excess is specific to centralised algorithmic assessment rather than a feature of hiring generally.

They also found disparities that only appear when each position is examined separately. Studying a vendor's data in aggregate, as earlier work did, washes them out. Position by position, using the standard set by US employment law, roughly a quarter of applications from Black applicants and 15 percent from Asian applicants went to positions that adversely affected them.

One structural caveat sits over all of it, and the authors state it themselves. They are the only group to have independently studied deployed hiring algorithms at scale, because vendors control the data. The finding has not been independently reproduced, and under current conditions almost nobody is in a position to try.

When everyone uses the same tool to judge people, the tool's blind spots stop being one company's mistake and become a wall that locks certain people out everywhere at once. Diversity of judgment is a feature, not a flaw: if different employers evaluate differently, a person rejected by one still has a real chance with another, and good people who do not fit one mold still find a place. The more hiring converges on identical automated filters, the more systematically it excludes whoever those filters happen to miss. If you use such tools, remember that everyone else may be using the same one, and the candidate it screens out may be screened out of the entire market, not just your role.

Source: Bommasani, Bana, Creel, Jurafsky and Liang, Algorithmic Monocultures in Hiring, ACM Conference on Fairness, Accountability and Transparency, 2026. One study, and the authors note they are the only group with the data access to conduct it.

The Alignment Problem

Brian Christian, 2020

On what happens when we hand consequential judgments to systems that share the same blind spots.