Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval

TL;DR

This paper audits bias in resume screening using a retrieval-based framework with Massive Text Embedding Models across nine occupations and over 500 resumes and job descriptions. It augments resumes with protected-name signals to test race, gender, and intersectional effects, revealing consistent favoritism toward White and male identities, and strongest bias against Black males in intersectional analyses. The study further shows that document length and name frequency modulate bias, and that these biases persist despite occupation patterns, underscoring the need for transparency, auditing, and mitigation in AI-driven hiring tools. Overall, the work provides a rigorous, scalable methodology for fairness assessment in resume retrieval systems and highlights critical policy implications for deploying LLM-based hiring technologies.

Abstract

Artificial intelligence (AI) hiring tools have revolutionized resume screening, and large language models (LLMs) have the potential to do the same. However, given the biases which are embedded within LLMs, it is unclear whether they can be used in this scenario without disadvantaging groups based on their protected attributes. In this work, we investigate the possibilities of using LLMs in a resume screening setting via a document retrieval framework that simulates job candidate selection. Using that framework, we then perform a resume audit study to determine whether a selection of Massive Text Embedding (MTE) models are biased in resume screening scenarios. We simulate this for nine occupations, using a collection of over 500 publicly available resumes and 500 job descriptions. We find that the MTEs are biased, significantly favoring White-associated names in 85.1\% of cases and female-associated names in only 11.1\% of cases, with a minority of cases showing no statistically significant differences. Further analyses show that Black males are disadvantaged in up to 100\% of cases, replicating real-world patterns of bias in employment settings, and validate three hypotheses of intersectionality. We also find an impact of document length as well as the corpus frequency of names in the selection of resumes. These findings have implications for widely used AI tools that are automating employment, fairness, and tech policy.

Figures (23)

• Figure 1: Relation of the three models investigated to a pre-trained LLM. Arrows between models indicate additional fine-tuning steps.
• Figure 2: Illustration of the resume screening as document retrieval framework. Task instructions are appended to job descriptions and treated as queries, while resumes are treated as documents. The cosine similarity between queries and documents estimates the relevance of a resume to a particular job description.
• Figure 3: For each occupation and model, cosine similarities are significantly higher (p$<$0.001) for resumes which belong to the same occupation as the job description (match) than those that belong to different occupations (unmatch), indicating the success of the document retrieval for resume screening framework.
• Figure 4: Resumes with White names are significantly preferred (p$<$0.05) in 85.1% of tests; those with Black names are preferred in 8.6% of tests. Gray regions indicate disparities which are not significantly different from zero (6.3% of tests).
• Figure 5: Resumes with male names are significantly preferred (p$<$0.05) in 51.9% of tests; those with female names are preferred in 11.1% of tests. Gray regions indicate disparities which are not significantly different from zero (37% of tests).