MIT Study: One Hiring Algorithm for All Firms Is Not Always Bad
MIT researchers argue that algorithmic monoculture, where many firms use the same algorithm to make decisions, is less harmful in hiring than past research suggested. In some cases it may even help job seekers, according to a study published in Philosophical Perspectives.
- MIT researchers say algorithmic monoculture in hiring is less problematic than past research suggested.
- They argue systematic exclusion is not compelling because the overall number of people hired is unaffected by firms sharing an algorithm.
- Monoculture can create informational echo chambers that hinder exploration, which the researchers prove mathematically.
- An ensemble algorithm combining several firms' algorithms can sometimes outperform using multiple algorithms, according to simulations.
- The study appears in Philosophical Perspectives; its practical feasibility remains to be explored.
Challenging a common worry about shared algorithms
AI tools increasingly replace human judgement in some settings. Resume screening algorithms, for example, are often used in hiring, where they may improve efficiency and consistency. Some scholars have raised concerns that this trend could end with one algorithm making all decisions in a given industry, a situation called algorithmic monoculture.
A frequent fear is systematic exclusion: a candidate rejected by one firm's algorithm would likely be rejected by every other firm's algorithm as well. The MIT team, led in discussion by co-authors Brian Hedden and Manish Raghavan, systematically evaluated this and other major objections. They concluded that many of these arguments either fail or are not decisive against all forms of monoculture.
Hedden is a professor in the Department of Linguistics and Philosophy with a shared position at the MIT Schwarzman College of Computing and the Department of Electrical Engineering and Computer Science (EECS). Raghavan is the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management and in EECS. Both are principal investigators in the Laboratory for Information and Decision Systems (LIDS).
"It depends on the details, like the domain we are talking about and the accuracy of the algorithm itself," Hedden says.
Exclusion, agency and gaming the system
Monoculture is not a new phenomenon. The researchers point to lending, where decisions were once made by independent bankers but now rely on standardized credit scores derived from the Fair Isaac Corporation (FICO) algorithm. A handful of resume screening algorithms are also commonly used by many Fortune 500 companies. The study focuses on hiring, though the approach could apply to other domains such as lending.
On systematic exclusion, the researchers used a series of models covering multiple situations. They argue the objection is not compelling because the overall number of people hired is not affected by firms using the same algorithm. Instead, monoculture could improve the bargaining power of candidates. "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," Raghavan says.
They also looked at objections about agency. A candidate whose resume is forwarded to every firm using the algorithm never gets a chance to adjust it. Hedden calls this a good objection to bad forms of monoculture, but says it does not hold up if candidates can revise and resubmit their materials.
Another concern is that people could game a single system, for instance by reformatting a resume. Hedden says it is not obvious that one algorithm would encourage this more than many different ones, since in the latter case a candidate might target a couple of firms' algorithms and gain an edge with a few employers.
Echo chambers and the ensemble fix
The researchers mathematically prove that monoculture tends to create informational echo chambers that can hinder exploration. This relates to the "wisdom of crowds" theory from social psychology, under which a diverse group of independent decision makers can outperform a single person. In hiring, diverse algorithms could produce a higher-quality pool of new hires, while a monoculture could cause candidates with the same characteristics and credentials to be hired every time by every firm. That may keep firms from finding better alternatives.
"Monoculture might inhibit the amount of discovery that happens overall. It is not clear if that is a bad thing, but it is definitely a worry when we think about designing AI for applications like science, art, or writing," Raghavan says. He adds that building randomness into a monocultural platform could induce more exploration.
Performance also matters. If a single algorithm is much more accurate than the many used by different firms, monoculture may be the better system. Another option is to package multiple firms' hiring algorithms into one "ensemble algorithm" that gives each candidate a score based on the average. In simulations of different hiring situations, the researchers confirmed that such an ensemble could sometimes outperform the use of multiple algorithms. Hedden says it remains to be explored how feasible ensembling would be in practice.
What it means for AI deployment
The findings suggest that the effects of shared algorithms cannot be judged in the abstract. The researchers note that other domains, such as generative AI content creation or AI-guided scientific research, may work differently, and monoculture there may be more problematic.
"A lot of the answers around the promises and pitfalls of algorithmic monoculture are going to be contextual," Raghavan says. He adds that more empirical work is needed. The researchers hope their study inspires further research on the long-term consequences of algorithmic monoculture.
