LLM-Based Resume Screening Shows Persistent Racial and Gender Bias

Researchers have uncovered significant bias in large language models (LLMs) deployed for resume screening, raising serious concerns about the fairness of AI-driven hiring tools now widely used by employers.

The Study Design#

The investigation tested a variety of Massive Text Embedding (MTE) models using a document retrieval framework that mimics real hiring decisions. The analysis examined resumes and job postings across nine different roles, drawing on a dataset of over 500 public resumes paired with 500 job descriptions. This setup allowed the researchers to measure how consistently the models ranked candidates with different demographic identifiers.

Key Findings on Bias#

The results paint a troubling picture of systematic discrimination. The embedding models exhibited substantial bias favoring applicants with White-associated names, selecting candidates with such identifiers in roughly 85% of screening scenarios. By contrast, candidates bearing female-associated names were favored in only about 11% of cases. The research also revealed an especially severe pattern for Black male applicants, who faced disadvantage in up to 100% of cases tested for certain job categories—a pattern that mirrors documented hiring discrimination in the real world.

Intersectional Effects#

Beyond simple race or gender analysis, the study confirmed that bias operates at the intersection of multiple demographic categories. Black applicants faced compounded disadvantages compared to their White counterparts, validating hypotheses about how intersectionality amplifies employment barriers. Additional factors—including the length of resumes and how frequently applicant names appear in training corpora—also influenced selection outcomes in measurable ways.

Implications for Industry and Policy#

These findings carry weight for HR departments, technology companies, and regulators. Automated hiring increasingly shapes who gets interviewed and hired, yet the models powering these tools appear to encode and reproduce historical patterns of employment discrimination. The research underscores the need for auditing AI systems before deployment and developing fairer alternatives to current resume-matching approaches.

For more details, see the full paper on arXiv.