MIT Study Finds Shared Hiring Algorithms Can Miss Better Candidates

If every employer uses the same hiring algorithm, a rejection at one firm may seem to predict rejection everywhere. A new MIT study challenges that as a blanket objection. Its sharper concern is that shared screening can keep firms looking at the same kinds of candidates, leaving better alternatives undiscovered.

The paper, “Algorithmic Monoculture and its Critics,” appears in Philosophical Perspectives. Researchers Brian Hedden and Manish Raghavan examined what happens when many employers rely on one algorithm rather than making decisions with different ones. Their work combines arguments about hiring, mathematical analysis and simulations; it is not a finding that any particular employer’s system has improved hiring.

A rejection is not the whole labor market

The familiar objection is easy to picture: a candidate screened out by one firm’s software encounters the same decision at every other firm using it. The researchers do not dispute that shared decisions can align. They argue that, in their hiring models, this does not reduce the total number of people hired. Employers still fill their jobs, but may compete for the same group of applicants.

That competition could strengthen those candidates’ bargaining power and push up their wages, Raghavan says. It is a narrower conclusion than saying no one loses out: an unchanged hiring total does not mean every applicant has the same chance of being selected.

Other objections also turn on design. If an application is sent to every firm at once, a candidate loses the chance to revise a résumé between attempts. Hedden says that criticism does not apply in the same way to a shared system that permits revisions and resubmission. Knowing how to format a résumé might also help applicants game one screen, though he argues that having multiple screens would not necessarily remove that incentive.

The candidates no one discovers

The study’s stronger warning concerns information. The researchers’ mathematical analysis finds that shared decisions tend to form informational echo chambers: employers repeatedly favor candidates with the same characteristics and credentials. When firms try different approaches, they may discover strong applicants that a common screen overlooks. In the shared-system case, the researchers say, even a filled position might not go to the best candidate.

That risk is not determined by the number of systems alone. A single, more accurate algorithm might make better choices than several weaker ones, the researchers say. Raghavan also suggests that adding randomness to a shared system could encourage more exploration. These are different ways to address the same problem: whether hiring software searches beyond its usual favorites.

Combining judgments is not the same as deploying them

The team tested that combined approach in simulations of hiring situations. Sometimes, an ensemble outperformed firms using separate algorithms. That result complicates the idea that employers must use different systems to benefit from multiple judgments: the judgments can instead feed one shared score.

Hedden says it remains unclear how feasible such an ensemble would be outside the simulations. The authors also caution that a job market has complexities their models cannot settle, and that the effects of shared algorithms may differ in areas such as AI-generated content or scientific research. Their finding is a reason to examine a system’s accuracy and room for exploration—not evidence that sharing one screen is always safe or harmful.