Apply to several companies that use the same hiring algorithm, and one unfavorable assessment could follow you across the job market. Employers could repeatedly overlook the same candidates because their software ranks people in similar ways.
New research from MIT challenges the idea that shared algorithms always produce worse hiring outcomes. In some cases, one shared algorithm can match or outperform a collection of different systems. The larger concern may be reduced exploration: when employers follow the same ranking, they may keep selecting from the same pool of candidates and miss people who could have performed well.
Why shared rankings don’t automatically mean fewer jobs
Algorithmic monoculture occurs when many organizations use the same algorithm to make similar decisions. In hiring, several employers might rely on one system to rank candidates before deciding whom to recruit.
In research highlighted by MIT on September 29, Brian Hedden and Manish Raghavan examine whether this kind of shared ranking necessarily harms applicants. Their paper, published in Philosophical Perspectives, uses theoretical models and simulations rather than employer data.
One part of their argument focuses on how hiring markets work. Candidates generally accept one job, while employers still need to fill their open positions. If several firms initially compete for the same highly ranked applicants, employers that fail to hire them eventually move farther down the candidate list. Under the model’s assumptions, the same number of people can still end up employed.
Competition for highly ranked applicants could also raise their bargaining power and wages. As Raghavan told MIT News, “the firms are fighting over the same pool of candidates.”
The finding applies to total employment in the researchers’ model. It does not show that applicants receive equal treatment or that a shared screening system cannot repeatedly disadvantage certain groups.
Shared rankings can limit what employers learn
Different hiring systems may rank the same applicant differently because each model uses different criteria or weighs information differently. Those differences can expose employers to candidates that a single shared ranking might consistently place lower.
This is one reason independent judgments can outperform a single assessment. The benefit depends on whether those judgments contribute different information rather than reproduce the same ranking.
Imagine that every employer favors applicants from familiar career paths. A candidate with an unconventional background might perform exceptionally well after being hired, but employers cannot learn from that outcome if every screening system rejects the person first. When many organizations rely on the same ranking logic, the same types of candidates can remain untested.
An earlier PNAS paper, co-authored by Raghavan, found that an algorithm that improves decisions for one employer can still reduce the quality of decisions across a hiring market. The newer research examines whether changes to the shared ranking process can preserve some of the information that separate systems would otherwise produce.
One option is an ensemble that combines several ranking models into one score. In the MIT simulations, ensembles sometimes matched or outperformed scenarios in which employers used separate algorithms.
The researchers also discussed randomness as a way to expose employers to candidates outside the highest-ranked group. These approaches remain theoretical. The authors acknowledge that the ensembles tested in their models may not translate to hiring systems.
Real hiring data shows how repeated rejection can happen
A separate Stanford-led study examined approximately 3 million applicants and 4 million applications screened by algorithms from one vendor.
The researchers found racial disparities affecting Asian and Black applicants. Among people who applied to 10 positions, 4% received rejection recommendations for every position, which was higher than researchers would expect by chance. The study measured algorithmic recommendations rather than confirmed final hiring decisions.
The MIT and Stanford studies address different questions. MIT examines whether shared algorithms necessarily worsen hiring outcomes under specific modeled assumptions. The Stanford-led research shows that repeated rejection patterns and disparities can appear when the same vendor’s screening system is used across applications.
A hiring market can therefore fill its vacancies while some applicants repeatedly fail to advance through screening. For companies evaluating these systems, overall hiring volume does not reveal whether the same people are consistently being filtered out.
Buyers may need to evaluate more than model accuracy
For enterprise AI buyers, the research suggests that accuracy benchmarks alone may miss an important part of the risk. If several employers rely on the same screening vendor or foundation model, similar ranking logic could cause them to overlook the same candidates.
Companies could examine whether different assessment methods produce substantially different candidate pools. They could also test whether applicants rejected by one screening method receive different results under another. This is an implication of the research rather than a process validated by the MIT study.
The authors also caution against applying the hiring findings to generative AI. Scientific research may depend heavily on exploring many possible ideas, while writing and art derive value from variation in outputs. The Neuron’s coverage of the “Artificial Hivemind” research examines a related issue: different language models can produce similar responses even when users expect independent outputs.
MIT’s findings depend on the accuracy of the shared algorithm and the environment in which organizations use it. For AI buyers, the question is therefore broader than whether one model performs well on its own. They may also need to examine what happens when many organizations rely on the same ranking logic.
