They Tried to Remove Gender from 709,000 Resumes. It Mostly Didn’t Work.
Degendering Resumes for Fair Algorithmic Resume Screening
Prasanna Parasurama, João Sedoc · arXiv:2112.08910 (PDF) · submitted December 16, 2021
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.
An obvious fix for gender bias in screening algorithms is to remove gender from the input. This study tests whether that’s actually possible, using a corpus of 709,000 resumes from IT firms — first training models to predict self-reported gender from resume text (measuring how much gendered signal exists), then iteratively removing it and watching what happens to screening performance.
The results are humbling for the "just anonymize it" position: resumes carry a significant amount of gendered information well beyond names and pronouns, lexicon-based scrubbing removes a lot of it — but only up to a point, and the residue persists in subtle word choices and structures.
What the paper reports
- Resume text contains substantial gendered information — enough to train classifiers that predict applicant gender.
- Lexicon-based obfuscation (removing gender-predictive tokens) reduces but cannot fully eliminate gendered signal.
- There is a measurable trade-off between gender obfuscation and screening-algorithm performance.
What this means for your resume
Our editorial interpretation — the paper does not give job-seeker advice.
- Blind screening is weaker in practice than employers may claim — "we remove names" removes only the loudest signal. This is worth knowing when evaluating claims that a hiring process is bias-free.
- The flip side: your writing style is identifiably yours. The controllable move is keeping content achievement-dense, since concrete accomplishments are the signal every screener — human or machine — is at least supposed to reward.
Read it with these caveats
arXiv preprint (v3, 2021); IT-industry resumes only. Bias-mitigation techniques have advanced since, but the core finding — gender pervades text beyond obvious tokens — is a property of language, not of any dated model.
Primary source: Degendering Resumes for Fair Algorithmic Resume Screening — 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.
Resume Search Engines Built on Embeddings Show Race and Gender Bias
Not the chatbots — the embedding models that power "find me candidates like this" search. Audited for gender, race, and intersectional bias.
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.