Fair AI in personnel selection

How to cite

Kandul, Serhiy; Leicht-Deobald, Ulrich. 2021. "Fair AI in personnel selection." AEA RCT Registry. AEARCTR-0007912. https://doi.org/10.1257/rct.7912-2.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Serhiy Kandul* , Ulrich Leicht-Deobald

* Primary investigator

Registry ID
AEARCTR-0007912
DOI
https://doi.org/10.1257/rct.7912-2.0
First registered
Last updated
Status
In development
Trial dates
2021-07-01 – 2021-07-10
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/7912

Abstract

Hiring algorithms are designed to help companies save time and resources in recruitment. Higher efficiency, however, might come at a cost of fairness, i.e. create disparities across social groups. In this experiment, we investigate how people perceive efficiency vs. fairness trade-off in personnel selection in a STEM context. We vary the degree of the tradeoff and explore the role of the (incentivized) beliefs on ground truth difference in qualifications across genders and on the backgroun inequalities of gender composition of he labour force.

Intervention

We present a scenario on hiring in the STEM context and introduce explore the efficiency vs. fairess tradeoff.
We explore two fairness metrics: satistical parity and equality of opportunities, and vary the degree of the tradeoff by looking at different levels of disparity across genders.

Experimental design

Participants face a scenario on hiring in the STEM context and then evaluate several screening algorithms which vary on the efficiency (operationalized as the number of evaluation rounds taken to fill a position) and parity (operationalized as acceptance rate of females and males candidates). We elicit participants' fairness perceptions of the algorithms and the choices of the algorithms.
We explore two fairness metrics (between-subject): satistical parity (disparities in acceptance rate across genders) and equality of opportunities (disparities in acceptance rate of qualified female and male candidataes).
We also vary the level of efficiency vs. fairness tradeoff (within-subject), with disparties in acceptance rate either low (0.97 ratio), medium (0.81, satisfying the 4/5 rule) or high (0.49 ratio). Conditions either favour female or male candidates (between-subject).
After the choices of the algorithms, we elicit (incentivized) beliefs on the ground truth differences in qualifications and background inqualities in gender composition of labour force in computer occupations. Finally, participants self-report general attitutes towards algorithmic and human decision making in HR context.

Design details

Randomization method
Randomization is done by the software (Qualtrics)
Randomization unit
individual
Clusters
280
Observations
280
Arms
70
Primary outcomes
The frequency of choices of the fair/unfair algorithms; fairness perceptions of the hiring algorithms; beliefs on cababilities of females and males
Secondary outcomes
Reasons behind choices (open text), control variables (general attitudes towards algorithm vs. human decision making)

Cited by

No citation data available for this registration yet.

Keywords