Everyone Using the Same Hiring Algorithm May Not Be the Problem, MIT Researchers Argue
Two MIT researchers argue that the main objections to every employer screening job applicants with the same algorithm do not hold up, and that averaging several firms' algorithms into one score can work as well as or better than letting each firm pick its own.
Brian Hedden, a professor in MIT's Department of Linguistics and Philosophy, and Manish Raghavan, of the MIT Sloan School of Management, weighed the case for and against algorithmic monoculture, one algorithm making the decisions across an entire field. Their paper, "Algorithmic Monoculture and its Critics," appeared online Sept. 25 in Philosophical Perspectives.
MIT's announcement of the work says the two went through the major objections and concluded that many either fail or are not decisive against every form of monoculture. One common objection is systematic exclusion: the worry that a candidate rejected by one firm's screening algorithm would be rejected by every firm using it. Working through a series of models, they argue it is not compelling because the number of people hired does not change when firms share an algorithm. "All the jobs get filled and the same number of people have jobs, but the firms are fighting over the same pool of candidates, which actually drives up wages," Raghavan says.
The researchers say they prove one cost mathematically: monoculture tends to create echo chambers of information that can hinder exploration, making it less likely that the best candidates get hired. Their answer is an "ensemble" that scores each applicant on the average of several firms' algorithms. Simulations showed it could sometimes beat a spread of separate ones. How feasible that would be in practice is untested, Hedden says.
The release notes that shared algorithms are not new: consumer lending already runs on credit scores derived from the Fair Isaac Corporation's algorithm, and a handful of resume screening algorithms are used across many Fortune 500 companies. The pair looked only at hiring, and say other uses, such as AI-guided scientific research, may work differently. "It depends on the details, like the domain we are talking about and the accuracy of the algorithm itself," Hedden says.
