The effects of an “algorithmic monoculture” depend on the details

MIT researchers argue that algorithmic monoculture — the use of one algorithm for all decisions in a domain — may not always be as harmful as some scholars have suggested. Brian Hedden, a professor in the Department of Linguistics and Philosophy with an MIT Schwarzman College of Computing shared position in EECS and a principal investigator in LIDS, and Manish Raghavan, the Drew Houston (2005) Career Development Professor at MIT Sloan and in EECS and a LIDS principal investigator, systematically evaluated major objections to monoculture. Their research appears in Philosophical Perspectives.

The researchers examined systematic exclusion, the concern that a candidate rejected by one firm's hiring algorithm would likely be rejected by every other firm's. Using models covering multiple situations, they argue this objection is not compelling because the total number of people hired is unaffected by firms sharing an algorithm. Raghavan said all jobs get filled and the same number of people have jobs, but firms compete over the same candidate pool, which drives up wages.

On agency, if candidates can revise and resubmit materials, the objection that they never get to adjust their resume does not hold, Hedden said. On gaming, Hedden said it is not obvious one algorithm would incentivize gaming more than many different firm algorithms.

The researchers mathematically prove monoculture tends to create informational echo chambers that can hinder exploration, making it less likely the best candidates get jobs in hiring. Bundling hiring algorithms into a single ensemble can overcome this, they show, sometimes performing as well as or better than polyculture. Simulations confirmed an ensemble algorithm could sometimes outperform multiple algorithms, though Hedden said feasibility in practice remains to be explored. They note domains such as generative AI content creation or AI-guided scientific research may work differently and monoculture there may be more problematic.