Job seekers are not automatically shut out when many firms share one hiring algorithm. On September 29, MIT News described a paper by Brian Hedden and Manish Raghavan. They call that shared use algorithmic monoculture and test the objections one by one in hiring. The paper appears in Philosophical Perspectives. This piece did not open the journal text.

[1]
Linocut: five identical profiles in a row, each with an ochre shadow behind it.
Five matching profiles stand in a row, each with an ochre echo. That is many firms sharing one algorithm, and the echo that follows. An illustration, not a hiring photograph., AI-generated illustration, not a news photograph

The usual worry is systematic exclusion: a resume screened out by one firm is screened out by every firm using the same algorithm. After checking a set of models, they argue that this is not compelling, because the number of people hired does not change just because firms share an algorithm. Raghavan says the jobs still get filled and the same number of people have jobs. Firms compete for the same candidates, and that drives wages up.

A candidate who cannot revise a resume and apply again is, Hedden says, a real objection to bad forms of monoculture. If the monoculture still allows revision and resubmission, that objection does not hold. The other side is gaming: a resume format that scores better can be copied. Hedden says it is not obvious that one algorithm encourages this more than many algorithms. With many, a candidate might target only a few firms.

[1]

What they prove is different. Monoculture tends to create informational echo chambers that hinder exploration. In hiring, that can make it less likely that the best candidates get jobs. Firms with different algorithms can, on the wisdom-of-crowds argument, hire a stronger pool. Monoculture can also hire people with the same traits every time, so firms discover fewer alternatives.

Bundling those hiring algorithms into one ensemble, and scoring each candidate on the average, can get past that limit. Simulations of hiring situations confirmed that an ensemble can sometimes beat the use of many separate algorithms, so monoculture can sometimes match or beat a polyculture. Hedden says it remains to be seen whether that ensembling is feasible in practice. If one algorithm is much more accurate than the many in use, monoculture may also be better. Randomness inside a single platform can increase exploration.

[1]

The result depends on the details. Hedden says a trend toward algorithmic monoculture is a realistic and important effect of AI, but it is hard to call it bad in the abstract. It depends on the domain and on how accurate the algorithm is. The study focuses on hiring, and it says generative-AI content and AI-guided scientific research may work differently. Monoculture in some of those domains may be more of a problem. Lending is treated as a related case, because bankers already share FICO credit scores. A handful of resume-screening algorithms are already common across Fortune 500 firms.

Raghavan says that, even as research, a lot of work remains on how to study these concerns empirically. The authors hope the paper leads to work on long-term consequences and on the real complications of a job market. The simulations are not observed wage or hiring statistics.

[1]

要点

  • In their models, one shared hiring algorithm does not reduce how many people are hired. Competition for the same people can raise wages.
  • The proved cost is an informational echo chamber and less exploration, so the best candidates are harder to find.
  • An ensemble of algorithms sometimes beat separate algorithms in simulation. The authors say practical feasibility is still open.
  • The focus is hiring. Generative content and AI-guided science are marked as possibly different.