Irrelevant Alternatives Bias Large Language Model Hiring Decisions

TL;DR

This work investigates whether large language models exhibit the attraction or decoy effect in AI-assisted hiring. Using a minimal, classical two-attribute design across six occupations, the authors prompt GPT-3.5 and GPT-4 in a recruiter role and compare target vs competitor choices with and without a decoy. They show consistent evidence of the attraction effect, with the decoy increasing the target's selection probability, and find that irrelevant attributes such as gender amplify the bias, especially for GPT-4 which also exhibits greater variance. Robustness checks with warnings and varied recruiter roles do not systematically remove the bias, highlighting the need for careful mitigation when deploying LLMs in high-stakes recruitment tasks and raising ethical considerations about context effects in AI-assisted decision making.

Abstract

We investigate whether LLMs display a well-known human cognitive bias, the attraction effect, in hiring decisions. The attraction effect occurs when the presence of an inferior candidate makes a superior candidate more appealing, increasing the likelihood of the superior candidate being chosen over a non-dominated competitor. Our study finds consistent and significant evidence of the attraction effect in GPT-3.5 and GPT-4 when they assume the role of a recruiter. Irrelevant attributes of the decoy, such as its gender, further amplify the observed bias. GPT-4 exhibits greater bias variation than GPT-3.5. Our findings remain robust even when warnings against the decoy effect are included and the recruiter role definition is varied.

Figures (10)

• Figure 1: Map of the decoy positions in a two-attribute space. Asymmetrically dominated (ASD-ed) decoys by the target are positioned in the green region. The brown region corresponds to phantom decoys, which are asymmetrically dominating (ASD-ing) the target. The map also shows the position of symmetrically dominated decoys and dominating phantom decoys by both the target and the competitor.
• Figure 2: Schematic overview of the method.
• Figure 3: An example prompt for the candidate selection task.
• Figure 4: Choice probabilities of the target candidate across 6 occupations in the control and treatment condition, and for two LLMs. The error bars represent the standard errors of the mean (SEM) over all six permutations of candidate presentation orders in the candidate selection prompt.
• Figure 5: Maps of the attraction effect bias on choices between target (T) and competitor (C) candidates, over bi-attribute job qualification space. Shown are results for GPT-3.5 (above) and GPT-4 (below), under six occupations and their required qualifications. The color intensity represents attraction effect strength, with redder shades indicating more positive bias and bluer shades representing more negative bias. Decoy candidates on the target-competitor line and left from it possess a valid working permit, while candidates to the right of the line are phantom decoys with no valid working permit.