Resume Search Engines Built on Embeddings Show Race and Gender Bias
Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval
Kyra Wilson, Aylin Caliskan · arXiv:2407.20371 (PDF) · submitted July 29, 2024
Summary by Andrew Johnson, OneTwo Resume editorial team · updated · We are not affiliated with the authors. arXiv papers are preprints and may not be peer-reviewed.
Most attention goes to chatbot-style screening, but a lot of real recruiting tech works differently: resumes are converted to embeddings — numerical representations — and retrieved by similarity to a job description. This University of Washington study audits that retrieval layer.
Using a document-retrieval framework that simulates candidate selection, the authors run a resume audit across Massive Text Embedding models, measuring whether resumes associated with different genders and races are retrieved at different rates for the same jobs — including intersectional combinations, which single-attribute audits miss.
What the paper reports
- The study builds a retrieval-based simulation of resume screening using MTE (embedding) models rather than generative chatbots.
- It audits selection rates across gender, race, and their intersections.
- It extends the audit-study tradition to the retrieval architectures that production candidate-search tools actually use.
What this means for your resume
Our editorial interpretation — the paper does not give job-seeker advice.
- The systems ranking you are often not chatbots — they’re similarity search. That makes the overlap between your resume’s vocabulary and the posting’s vocabulary mechanically decisive, which is the honest case for tailoring each application.
- The de-identification advice (minimal demographic signals) applies to this architecture too.
Read it with these caveats
arXiv preprint (v3). Results are for specific open embedding models; commercial vendors’ proprietary models may differ. Retrieval-rate differences in simulation are strong evidence of encoded bias but not a measurement of any single employer’s pipeline.
Primary source: Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval — always read the paper before citing it. Spotted an error in our summary? Tell us and we'll fix it with a visible correction.
More in bias & fairness in ai screening
The Famous Name-Bias Experiment, Rerun on ChatGPT-Era Models
The landmark 2003 field experiment — identical resumes, racially suggestive names — replicated against GPT-3.5, Bard, and Claude.
Newer LLMs Show Less Name Bias — But Judge Your University Instead
A 2025 audit finds explicit gender/race bias has receded in recent LLMs — while implicit bias around educational background remains significant.
Overcorrection Is Real: Benchmarking Reverse Gender Bias in LLM Resume Scoring
A benchmark grounded in labor economics finds significant reverse gender bias and over-debiasing in LLM resume scoring.