Krebsbach, Jennifer M., Jane E. Lee, Steven Zeck, Arti Thakur, and Martin Hilbert. 2026. “Does AI Reinforce Gender Stereotypes? Gender Bias in Job Descriptions Generated by 40+ Large Language Models.” Gender, Work & Organization. https://doi.org/10.1111/gwao.70291.

Abstract: Human resource (HR) professionals and office staff are often responsible for competing tasks that quickly accumulate. To increase efficiency, some HR staff are turning to large language models (LLMs) to automate tasks, including creating job postings based on job titles alone. We sought to examine how this practice has the potential to influence applicant pools and hiring outcomes, with attention to gendered effects in organizational settings. Using a broad sample of 48 LLMs, we conducted an algorithm audit to evaluate whether the use of a gender-neutral or gender-coded job title would influence the output created by LLMs. We measured salary disparities, unmasked comparative narrative ratings, and BSRI-informed discursive patterning utilized throughout the job description. We selected 32 jobs and created a female-coded, male-coded, and gender-neutral version of each job title (e.g., firewoman, fireman, and firefighter), studying a total of 4608 job descriptions. In comparing the LLM outputs, we found that for white-collar job titles, male-coded job descriptions were assigned higher salaries, but this was not robust across all analyses. Across occupations, advertisements generated from gender-coded titles received higher unmasked comparative LLM ratings than matched advertisements generated from gender-neutral titles, with female-coded advertisements receiving the highest ratings. Minimal optimization appeared to substantially reduce these ratings and perform comparably to maximal optimization, but neither reached the lower ratings observed when generation began with a gender-neutral title. As the use of AI tools is becoming commonplace, our findings emphasize the need for feminist, organizationally accountable approaches to AI-assisted HR work. Simple prompt modifications can mitigate, but not fully eliminate, some measured gendered differences in LLM-generated job descriptions, but they do not eliminate deeper occupationally gendered patterns.

Corresponding Author: Jennifer M. Krebsbach is a PhD candidate at the University of California Davis. Her research focuses on identifying barriers and biases in organizational structures that impact occupational mobility for women and multiply marginalized populations using mixed methods designs.