MIT researchers Brian Hedden and Manish Raghavan have examined the consequences of employers relying on the same hiring algorithm. Their theoretical models and simulations suggest that familiar objections do not settle whether every form of algorithmic monoculture is harmful. Outcomes depend on the system’s accuracy and the setting in which decisions occur.
A more specific concern is lost discovery: organizations drawing on similar information may repeatedly select similar candidates and overlook alternatives. Combining multiple algorithms into an ensemble sometimes improved results in the simulations. The researchers say practical feasibility and real-world consequences still need investigation; the work does not establish that shared screening tools are fair or effective in actual hiring.
NAOAT analysis: Look beyond the number of vendors
For an employer, an evaluation should ask what information the screening process can miss. Testing only whether a tool agrees with past hiring decisions risks making agreement the objective. A useful review would also examine applicants who would have been excluded and document whether their skills were assessed by an independent route.
Buying two products is not necessarily the same as obtaining two independent perspectives. Procurement teams should ask about shared models, training sources and screening criteria, while recognizing that suppliers may disclose only part of that information. Uncertainty about overlap belongs in the evaluation record.
An ensemble also needs an explanation of how disagreements are resolved. Averaging scores might hide a strong signal that appears in only one assessment. Teams should define which evidence can trigger human review and how applicants can correct inaccurate information.
The practical question is therefore which candidates the process gives the organization a chance to discover. Before changing hiring policy, an employer should compare actual outcomes over time and examine the consequences of errors, rather than treating a simulated improvement as a deployment decision.