Recent analyses of large language models (LLMs) used for resume screening reveal significant biases, consistently favoring candidates with names perceived as white and male. This pattern holds across multiple studies, suggesting a deeply embedded issue in how these AI tools process applicant information. The core finding is that AI systems, when tasked with initial resume evaluations, demonstrably prefer certain demographic markers over others, specifically against names associated with Black individuals and women.
Studies, including one published on January 22, 2025, indicate that nearly 90 percent of the time, AI models favored resumes with male-sounding names. Another investigation, released October 31, 2024, corroborated these findings, highlighting gender, racial, and intersectional biases in three state-of-the-art LLMs. Researchers systematically altered first names on resumes to represent typically white and Black men and women, observing distinct ranking disparities. Notably, resumes associated with Black males were never preferred over those with white male names. The smallest preference difference was observed between typically white female and white male names.
Read More: AI Training Uses Independent Content Without Paying Publishers
Methodical Investigations Uncover Pattern
A project exploring 'hiring-bias' on GitHub outlines a process for counterfactual audits of LLM resume screening. This involves generating resume variants by altering just one demographic signal to measure how verdicts change. The methodology is designed to isolate and quantify bias by comparing outputs when only a single factor, such as a name associated with a particular race or gender, is modified.
Several research groups have focused on specific LLM providers. A University of Washington study, detailed in reports from October and November 2024, utilized models from Mistral AI, Salesforce, and Contextual AI. These models were tested against over 500 real-world job listings using a dataset of over 550 real-world resumes. Another analysis focused on open-source LLMs from Salesforce, Contextual AI, and Mistral.
Broader Implications and Regulatory Shifts
The findings have broader implications as businesses increasingly integrate digital tools into their hiring workflows. This trend has prompted regulatory action, with New York City enacting a law that requires companies using AI hiring systems to disclose their performance metrics.
Read More: UK Govt Offers Army Jobs to Unemployed Youth Amid Record Joblessness
These investigations build upon ongoing work to understand and mitigate biases within AI systems. Research published on March 24, 2025, delves into intersectional bias in causal language models and auditable LLMs for race and gender bias, questioning AI's capacity to replace human judgment in these sensitive areas. The underlying concern is that these AI tools, despite their technological sophistication, are not neutral arbiters but reflect and potentially amplify existing societal prejudices.
