Audit Studies of LLM and Resume-Screening Algorithms

What the evidence says

Audit studies of LLM-based resume screening find race and gender differences in how identical resumes are ranked. In a retrieval test of three massive text embedding (MTE) models over more than 3 million comparisons, resumes with White-associated names were preferred in 85.1% of cases and resumes with female-associated names in 11.1%. A minority of cases showed no difference. White male names were preferred over Black male names in all 27 tests (p<0.05). Male names were preferred over female names in 51.9% of 27 gender tests, and female names in 11.1%. Race preference was therefore stronger than gender preference in this setting. Black female names were preferred over Black male names in 66.7% of tests, while White female names were preferred over Black female names in 48.1% of tests.

The size and direction of these gaps depend on low-level text features. Title-only resumes showed differences in 96.2% of 27 race tests, against 93.7% for full-length resumes. When names were matched to roughly equal frequencies in the training corpus, the race pattern reversed. Black names were preferred in 51.9% of tests and White names in 22.2%. The sources point to name frequency as one pathway for the measured preference.

Generative and ranking audits of GPT-4 find similar gaps. In a resume-scoring task, White names scored 1.14 points above Asian names on a 1-100 scale (p=0.006). Women scored 0.28 points below men (p=0.052). When the model wrote resumes, women's resumes had 16.26 fewer months of experience on average (p=0.022). Asian names' resumes listed language skills 45% of the time, and none of the White or Black names' resumes included non-U.S. experience, against 15 of 80 Asian names and 3 of 80 Hispanic names. No Black names were assigned engineering occupations. The authors report that resume generation showed more and larger biases than resume assessment.

For disability, GPT-4 ranked a disability-enhanced resume first less often than a control resume. The rate varied by disability: 0 of 10 trials for Autism, 1 of 10 for Deaf, 2 of 10 each for Depression and Cerebral Palsy, and 5 of 10 each for Blindness and general disability. A GPT-4 variant given disability-justice and DEI guidance ranked the enhanced resume first in 37 of 70 trials (53%), against 15 of 70 (21%) for unmodified GPT-4. It did not change the rate for Depression (2 of 10 trials in both conditions). Text analysis found GPT-4 over-used the word "DEI" when describing enhanced resumes.

Model quality also varies. In a simulation, Claude-Sonnet-4 selected the more-qualified candidate in 96% of pairs, and Llama-3.1-8B in 74%. For equally qualified pairs, most models had discriminant validity below 0.9 when demographic information was present. Smaller open models did no better than random guessing. Criterion validity fell by 0.09 on real Reddit resumes. In a separate counterfactual audit of a purpose-built classifier, gender-associated name swaps changed the top role in 14.76% of comparisons, race-associated swaps in 10.43%, and age cues in 7.72%. A model with those cues removed did not change at all.

Evidence on human reviewers working with AI recommendations is thinner and associational. Not having AI recommendations was associated with 55.6% longer resume viewing. When AI recommendations were extremely severe and incongruent with stereotypes, non-white candidates' resumes were viewed 20.1% less time than white candidates' for high-status occupations. Completing an implicit association test first was associated with no measured difference in viewing time between Black and white candidates for high-status occupations.

Applicants also manipulate screening systems. About 1% of 196,682 resumes contained hidden prompt injections. The share rose to about 1.2% in 2024, from 0.6% to 0.8% in 2019-H2 through 2023-H2. Over 90% used data injection rather than explicit instructions, and none used optimization-based attacks. General-purpose detectors had 7.0% and 5.0% recall on this data. No source on this page measures whether these injections change hiring outcomes.

Where the evidence splits

The gender gap is not consistent across studies. The MTE retrieval audit finds male names preferred in 51.9% of tests, while the GPT-4 scoring audit finds women rated 0.28 points lower at p=0.052, which is just above the 0.05 threshold. The authors of the second study note that a different prompt showed no gender effect. The sources do not explain the difference, though the studies use different tasks (retrieval ranking versus 1-100 scoring) and different models.

The direction of the race gap depends on name frequency. With population-proportional frequencies, White names were preferred. With names matched to equal corpus frequencies, Black names were preferred in 51.9% of tests. The study itself ties the reversal to name frequency, so this is a split within one design. The sources do not say which setting resembles deployed systems.

The link between viewing time and selection depends on whether the AI recommended the candidate. For candidates not recommended by AI, selection chance rose 3-4% for every additional 30 seconds of viewing. For recommended candidates, duration had a negative association with selection. The sources do not explain the difference.

Thresholds and dose-response

• Qualification gap: many models had criterion validity below 0.90 when candidates differed by one relevant qualification (k=1). Validity rose at k=2 and k=3, and some models exceeded 0.95 at k=3.
• Selective review: a threshold set for 80% calibration coverage produced 0.8130 test coverage, 0.2955 accepted-case accuracy (up from 0.2489), and a 0.1870 review rate. It captured 23.73% of all errors.
• Resume length: title-only resumes showed race differences in 96.2% of tests, against 93.7% for full-length resumes.
• Name frequency: a preference for White names under population-proportional frequencies reversed to a preference for Black names under equal corpus frequencies.

Open empirical questions

• Do the preferences found in embedding models and GPT-4 audits appear in the commercial resume screeners that employers actually use?
• Do hidden prompt injections change screening outcomes, and do they change them differently for different demographic groups?
• What size does the name-frequency reversal take across other model families, name sets and occupations?
• Do the disability-ranking gaps and the effect of disability-aware instructions hold with more than 10 trials per condition and for conditions beyond those tested?