MIT study finds shared hiring algorithms can help applicants under some conditions

MIT researchers say companies using the same hiring algorithm could, in some circumstances, benefit job seekers rather than shut them out. In a study highlighted by MIT on September 29, Brian Hedden and Manish Raghavan argue that the outcome depends on the algorithm’s accuracy and design. Their comparisons are based on models and simulations, so they do not show that applicants have gained jobs or higher wages in real hiring markets.

The question matters as employers use automated tools to assess applications. If several firms rely on one system, an applicant ranked poorly by that system may face similar assessments at each firm. Hedden and Raghavan call this arrangement ‘algorithmic monoculture’ and compare it with a market in which employers use different algorithms. Their work examines what each arrangement might mean for applicants and for the information employers gain from their decisions.

Why a shared system might help some candidates

One concern is that a shared algorithm will repeatedly exclude the same people. Hedden and Raghavan argue that, in their hiring models, using the same algorithm does not itself change the number of jobs filled. That does not mean every applicant has the same prospects: those ranked poorly may fare badly across multiple employers. The authors’ point is narrower. Similar rankings do not, by themselves, establish that fewer people are hired overall.

The researchers also identify a possible advantage for candidates near the top of a shared ranking. If firms seek the same pool of highly ranked applicants, they may have to compete for them. Raghavan says that competition could push up those candidates’ wages. This is a proposed effect of the model, not evidence that wages have risen where employers use common screening software.

The cost of making similar choices

The authors find a stronger case against a single ordinary algorithm in what employers might fail to learn. If firms repeatedly favor familiar types of candidate, they may explore fewer alternatives and miss people who would perform well. Different algorithms can draw attention to different applicants. That diversity may improve the information available across employers over time, even where the shared system appears consistent in individual decisions.

Hedden and Raghavan propose one possible way around that problem: a shared ‘ensemble’ that combines the assessments of multiple algorithms. In principle, this could preserve information from different approaches while giving firms a common ranking. In the paper’s simulated comparisons, that ensemble performed better than firms using separate algorithms, which in turn performed better than an ordinary shared algorithm. The result depends on the assumptions used to model candidate value and imperfect estimates of it.

The simulations used 1,000 candidates per run and averaged results across 1,000 runs. They compared modeled hiring arrangements, not employment records. Hedden says it remains unclear whether employers could put the proposed ensemble into practice. The paper also leaves open how its conclusions would change when actual firms, applicants and screening systems interact outside the model.

What deployed screening systems show

Separate research offers a view of a narrower real-world problem. Researchers affiliated with Stanford, Chapman and Northeastern universities examined one vendor’s screening recommendations for 4,197,168 applications from 3,372,132 applicants, covering 1,746 positions at 156 employers. Their data span December 2018 to December 2022. According to the study, the vendor supplied the data but could not edit or veto the researchers’ analysis.

Among applicants who applied to ten positions, 4% received rejection recommendations for all ten, more than the researchers’ statistically independent baseline predicted. The study also found that 25.87% of applications from Black applicants and 14.74% of applications from Asian applicants went to positions where their groups experienced adverse impact under its position-level measure. These figures describe screening recommendations and a statistical measure; they do not establish final hiring outcomes or a legal finding of discrimination.

That study shows how correlated recommendations can arise in a deployed system, but it covers one vendor and does not test the ensemble proposed by Hedden and Raghavan. It cannot settle whether a differently designed shared algorithm would produce the benefits MIT’s researchers describe. Equally, the MIT models do not remove the concern raised by the observed rejection patterns. The two studies address different parts of the question: possible outcomes under specified conditions and recommendations recorded in one system.

Hedden and Raghavan say the effects of a shared algorithm depend on the domain as well as the system’s design and accuracy. They focus chiefly on hiring and caution that uses such as generative AI or scientific research may behave differently. For employers and applicants, the unresolved issue is how any proposed shared system performs in practice, including whether it broadens opportunity while avoiding repeated poor assessments of the same candidates.