Science1 publisher2 min readPublished
MIT researchers tie shared hiring algorithms to echo chambers that curb exploration
MIT's Brian Hedden and Manish Raghavan argue that the provable cost of every firm using one hiring algorithm is an echo chamber that curbs exploration. In their models, an ensemble of several screening tools can offset that loss.
The Scientist · Science desk
What happened
- The pair tested the systematic-exclusion objection across a series of models and found it unconvincing, because the total number hired does not change when firms share one algorithm.
- They reached the same verdict on many other objections to monoculture, concluding that each either fails or is not decisive against every form of it.
- A handful of resume-screening algorithms are already in common use across many Fortune 500 companies, according to the MIT account.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- decision Employers licensing a common screener face a design choice the models say matters: a single algorithm, or an ensemble of several that can offset the lost exploration.
- exposure Because the rebuttal is argued from total hires and wages, an applicant the shared screen marks down still meets the same verdict at every firm that uses it.
- constraint The reassurance covers hiring and perhaps lending; the authors say monoculture in generative AI content or AI-guided science may be more of a problem.
The evidence is theoretical. Brian Hedden, a professor in MIT's Department of Linguistics and Philosophy, and Manish Raghavan of MIT Sloan and electrical engineering built a series of models covering multiple situations and proved results inside them. The paper appears in Philosophical Perspectives [7][11]. A proof tells you what must follow once the assumptions are fixed. The MIT summary does not set out those assumptions or report an effect size from any real hiring market.
The exclusion objection is about particular people: the applicant one firm's algorithm screens out is likely to be screened out by all of them [2]. The rebuttal counts total hires [11], and Raghavan took the point on to pay. "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," he said [13]. The researchers argue that candidates' bargaining power could rise as a result [12].
Where the authors do prove a cost, it concerns information. Monoculture, they show, tends to create informational echo chambers that can hinder exploration. In hiring, that could make it less likely that the best candidates get jobs [4]. The hedges in that finding are the authors' own. Their fix, bundling several hiring algorithms into one ensemble, can overcome the limitation, and could sometimes let a monoculture perform as well as or better than a polyculture, where firms use different algorithms [5].
Raghavan worries about scale. "The worry is that, as more people use AI and algorithms to get information and make decisions, there is more of a vehicle for this kind of correlation to occur," he said [9]. Credit offers an older case. Bankers who once judged borrowers independently now all work from standardized scores derived from the FICO algorithm [8].
Hedden said it is hard to judge monoculture in the abstract. "It depends on the details, like the domain we are talking about and the accuracy of the algorithm itself," he said [6]. For anyone licensing a screening tool, I think the ensemble result is the most useful finding, as long as Hedden's accuracy condition comes with it [5][6].
What to watch
- Whether the paper's models specify when an ensemble matches or beats a polyculture, and how accurate its component algorithms must be.
- Any empirical test in markets dominated by a few resume screeners showing whether exploration, and the hiring of the best candidates, actually falls.
- How the authors treat the agency objections and non-hiring domains such as AI-guided research, where they expect monoculture may be worse.