How shared hiring algorithms can miss better candidates

An MIT theoretical study separates repeated rejection from reduced discovery and finds that averaging several algorithms can sometimes improve simulated hiring results.

An MIT theoretical study separates repeated rejection from reduced discovery and finds that averaging several algorithms can sometimes improve simulated hiring results.
Employers using the same hiring algorithm may repeatedly select candidates with similar characteristics and credentials, leaving potentially better alternatives undiscovered. But a shared system need not perform worse than separate systems: mathematical models and simulations by MIT researchers Brian Hedden and Manish Raghavan suggest that its accuracy and design can change the result.
MIT News described the research on September 29, 2026. The study, which appears in Philosophical Perspectives, examines “algorithmic monoculture”—using the same algorithm, or computer procedure for making decisions, across a particular field. In hiring, the comparison is with employers using different algorithms to screen applicants.
Analysis
One concern is systematic exclusion: an applicant rejected by one employer’s algorithm may also be rejected by every other employer using it. The decisions are correlated, meaning they tend to agree rather than give the applicant separate chances.
Hedden and Raghavan argue that this objection is not decisive in their hiring models because using a shared algorithm does not change the total number of people hired. That result concerns the number of hires; it does not establish that exclusion is harmless or that opportunities are fairly distributed. Raghavan also says employers competing for the same pool of candidates could strengthen those candidates’ bargaining power and drive up wages.
The discovery problem asks a different question: are employers finding the best candidates? According to MIT News, the researchers mathematically show that shared algorithms can hinder exploration. Different algorithms may identify different promising applicants, while one shared system may repeatedly favor the same kinds of candidates. Filling every job therefore does not necessarily mean finding the strongest available hires.
Analysis
Diversity of algorithms is not automatically better. In the researchers’ analysis, a shared algorithm may perform better if it is much more accurate than the separate algorithms employers would otherwise use. Raghavan also suggests adding randomness to a shared platform to increase exploration and reduce the risk of overlooking alternatives.
Another option is an “ensemble algorithm”: a single system that combines several employers’ hiring algorithms by averaging their scores for each candidate. This retains multiple assessments inside one shared decision process. MIT News reports that Hedden and Raghavan’s simulations found such a system could sometimes outperform employers using separate algorithms—not that averaging always produces better results.
These findings come from mathematics and simulated hiring situations, not measured outcomes in real hiring markets. Hedden says the feasibility of combining algorithms in practice remains to be explored, and Raghavan says further research using real-world evidence is needed to understand the consequences.

The macOS and Windows preview lets Copilot use on-screen controls. Removing a saved approval does not revoke access already granted in a running session.
3 min

Nordic Makers invested as co-founder Lasse Finderup took majority ownership. The new tool connects recordings with earlier interview material, according to the company.

The company says at least 26,000 SI agents scanned more than 25,000 assets for routes to a breach. Verification details and comparisons with other testing methods remain unclear.
2 min
2 min