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

MIT researchers Brian Hedden and Manish Raghavan have published a paper in Philosophical Perspectives arguing that algorithmic monoculture — the situation in which a single algorithm makes all decisions across a domain — is not as uniformly harmful as some scholars have suggested. Hedden is a professor in the Department of Linguistics and Philosophy and holds a shared position with the Department of Electrical Engineering and Computer Science (EECS) at the MIT Schwarzman College of Computing, and is a principal investigator in the Laboratory for Information and Decision Systems (LIDS). Raghavan is the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management, also holds a shared position in EECS, and is a LIDS principal investigator. The two systematically evaluated the main objections raised against monoculture, focused their analysis on hiring, and concluded that several of the leading arguments either fail outright or are not decisive against all forms of monoculture.
The paper's central positive claim is a mathematical proof that monoculture tends to create informational echo chambers that hinder exploration. In a hiring context, the researchers say this could make it less likely that the best candidates end up in jobs. Their proposed mitigation is to bundle multiple firms' hiring algorithms into a single ensemble algorithm, which could assign each candidate a score based on an average. Through simulations of different hiring situations, they confirmed that such an ensemble can sometimes outperform the use of multiple distinct algorithms, allowing monoculture to perform as well as, or better than, a polyculture in which different firms use different algorithms. Hedden noted in a statement that a trend toward algorithmic monoculture is a realistic scenario and an important issue brought about by AI, but that whether it is bad depends on the details, such as the domain and the accuracy of the algorithm itself.
Algorithmic monoculture is not new. Lending decisions that were once made independently by bankers at individual banks now rely on standardized credit scores derived from the Fair Isaac Corporation (FICO) algorithm, and a handful of resume screening algorithms are commonly used across many Fortune 500 companies. Raghavan said the worry is that as more people use AI and algorithms to gather information and make decisions, there is more of a vehicle for this kind of correlation to occur. The researchers focused on hiring but say the approach could extend to lending, while noting that domains like generative AI content creation or AI-guided scientific research may behave differently and that monoculture there may be more problematic.
The first objection they examined was systematic exclusion: the idea that if one firm's algorithm screens out a candidate, every other firm using the same algorithm would too. After evaluating this using a series of models capturing multiple situations, they argue the objection is not compelling because the total number of people hired is unaffected by firms sharing an algorithm. Raghavan said all jobs still get filled and the same number of people have jobs, but firms end up competing over the same candidate pool, which drives up wages. This reframes monoculture as a potential boost to candidate bargaining power rather than a mechanism of exclusion in the aggregate.
A second family of objections concerns agency. If a candidate's resume is forwarded to every firm running the same algorithm, the candidate never gets a chance to revise materials to improve their odds. Hedden said this seems like a good objection to bad forms of monoculture, but it does not hold up if the monoculture allows candidates to revise their resume and resubmit their materials. The flip side is gaming: if a particular resume format produces better outcomes, candidates could simply reformat. Hedden argued it is not obvious that a single algorithm incentivizes more gaming than many different algorithms, since in the polyculture scenario a candidate might target a couple of firms' algorithms to gain an advantage with a few employers.
The researchers also considered a less-explored objection: homogenization of information. Drawing on the wisdom of crowds, a theory from social psychology, Hedden explained that a diverse group of independent decision makers can outperform a single decision maker — so in hiring, firms with diverse algorithms could produce a higher-quality pool of new hires. Monoculture could cause candidates with the same characteristics and credentials to be hired every time by every firm, potentially preventing firms from discovering better alternatives. Raghavan said monoculture might inhibit overall discovery, which is not clearly bad but is definitely a worry when designing AI for applications like science, art, or writing. He added that building randomness into a monocultural platform could induce more exploration.
Performance also depends on the algorithm itself. If one algorithm is substantially more accurate than the many algorithms different firms would otherwise use, monoculture may be the better system. Ensembling — packaging multiple firms' hiring algorithms so a candidate receives an average score — is a concrete way to capture that gain, and the simulations confirmed it can sometimes beat using multiple algorithms. Hedden said it remains to be explored how feasible such ensembling would be in practice. Raghavan added that answers about the promises and pitfalls of monoculture will be contextual, and that from a research perspective much work remains to approach these concerns empirically. The authors hope the work spurs further study of the long-term consequences of monoculture and of the real-world complexity of systems like a job market.
Why it matters: Teams building or procuring screening, ranking or lending models now have a research-backed framing in which the strongest documented downside of shared algorithms is reduced exploration rather than aggregate exclusion — an inference from the authors' argument, not a claim the paper makes about any specific product. For developers, that shifts the design question toward diversity, randomization and ensembling as mitigations, while the paper's own limits suggest the answer will remain domain-specific and, on ensembling, not yet shown to be practical.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
In a study focusing on hiring decisions, MIT researchers found the use of a single algorithm by many firms could benefit job seekers in certain situations.