How “algorithmic monoculture” plays out hinges on the specifics

Meta description: MIT researchers argue algorithmic monoculture in hiring may not mean exclusion, but echo chambers, wage effects and ensembles shape outcomes.

Tags: algorithmic monoculture, AI hiring, MIT research, ensemble algorithms

Featured image: Abstract friendly illustration of identical algorithmic gears forming one hiring funnel, no people, no logos, no text.

For anyone involved in hiring, job seeking, or AI governance, a new MIT analysis offers a more careful view of algorithmic monoculture, the case where one algorithm guides decisions across a domain. AI tools are taking over some human judgments. Resume screening systems are common in hiring and can make decisions faster and more consistent. Some scholars fear automated systems could converge on one algorithm that shapes every decision in an industry, with negative consequences for so-called algorithmic monoculture. The researchers tested the main objections, including systematic exclusion, and concluded that they fail or are not decisive against every form of monoculture. Instead, their mathematical proof shows that monoculture tends to create informational echo chambers that limit exploration. In hiring, that can reduce the chance that the best candidates get jobs. But combining several hiring algorithms into one ensemble can address that problem, the study says. In some cases, monoculture may match or beat polyculture, where firms use different algorithms.

The study appears in Philosophical Perspectives. Brian Hedden, a professor in MIT’s Department of Linguistics and Philosophy, holds an MIT Schwarzman College of Computing shared position with the Department of Electrical Engineering and Computer Science (EECS) and is a principal investigator in the Laboratory for Information and Decision Systems (LIDS). He co-authored the paper with Manish Raghavan, the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management and in EECS, also a principal investigator in LIDS.

How a single algorithm can spread

The concern is not only theoretical. Independent bankers at individual banks once made lending decisions, but today many use the same information based on a borrower’s standardized credit scores, which come from the Fair Isaac Corporation (FICO) algorithm. In hiring, a small number of resume screening algorithms are already common across many Fortune 500 companies. Raghavan says the worry grows as more people use AI and algorithms to get information and make decisions, because that creates more scope for correlated outcomes.

The authors focused on hiring, but their method could apply to other domains such as lending. They note, however, that areas such as generative AI content creation or AI-guided scientific research may work differently, and monoculture in some of those fields could be more problematic. Hedden says a trend toward algorithmic monoculture is realistic and important because of AI’s spread, but it is hard to judge in the abstract. The outcome depends on the domain and the accuracy of the algorithm itself.

Why systematic exclusion may not be the central threat

One common objection is that the same algorithm used by many firms could create systematic exclusion. If one company screens out a candidate’s resume, that person may face the same result at every other company using the same system. Hedden and Raghavan evaluated this argument using a range of models designed to reflect multiple situations. They conclude it is not compelling because the total number of people hired does not necessarily change when firms use the same algorithm.

The study instead suggests monoculture could improve job candidates’ bargaining power. If all jobs are filled and the same number of people have jobs, but firms compete over one shared candidate pool, wages may rise. The researchers also looked at objections tied to agency. If a candidate’s resume is automatically forwarded to every firm using the same hiring algorithm, the candidate may never get a chance to revise it. Hedden says that is a strong objection to bad forms of monoculture, but it weakens if the system allows candidates to revise and resubmit their materials.

There is also the risk of gaming. If a particular resume format produces a better outcome, candidates could reformat their materials to improve their chances. But the authors say it is not obvious that one algorithm encourages gaming more than many algorithms used by different firms. In a mixed system, a candidate might target a few firms’ algorithms and try to game them, gaining an advantage with only some employers.

The real issue: echo chambers and less discovery

The less-explored objection is that monoculture can homogenize information. The study connects this to the wisdom of crowds, a social psychology theory saying a diverse group of independent decision makers can outperform a single person. In hiring, that means firms using diverse algorithms may assemble a higher-quality pool of new hires. Monoculture could cause candidates with the same characteristics and credentials to be hired repeatedly by every firm, which may stop firms from discovering better alternatives.

Raghavan says monoculture might reduce overall discovery. That may not be automatically bad, but it is a concern when designing AI for applications such as science, art, or writing. The researchers say the problem can be softened by building randomness into a monocultural platform to encourage more exploration. They also note that monoculture’s performance depends on the algorithm. If one algorithm is much more accurate than the many algorithms used by different firms, monoculture may be the better system.

One practical idea is an ensemble algorithm. It packages several firms’ hiring algorithms into one system and assigns each candidate a score based on the average. The researchers ran a series of simulations of different hiring situations and confirmed that this ensemble approach could sometimes outperform the use of multiple separate algorithms. Still, they say it remains unclear how feasible algorithmic ensembling would be in practice.

What this means for readers

The authors stress that the promises and pitfalls of algorithmic monoculture are contextual. Even as a research agenda, they say much more empirical work is needed to understand how real systems create or reduce these concerns. Their hope is that the study encourages research into the long-term consequences of algorithmic monoculture and into the real-world complexities of a system as broad as the job market.

Who should use it: HR teams, platform designers, researchers, and policy analysts who need a balanced framework for thinking about algorithmic monoculture, hiring accuracy, and ensemble design. Who should skip it: readers looking for a simple verdict that AI hiring systems are either safe or dangerous, because the study’s answer depends on the domain, algorithm accuracy, and whether the system allows revision, randomness, or shared scoring.