# PubMed Central: An and colleagues find intersecting biases in LLM resume scores

[Canonical reading](https://escapemeridian.com/readings/pubmed-central-an-and-colleagues-find-intersecting-biases-in-03834176)

Jiafu An and colleagues report that five large language models gave different resume scores to fictitious applicants whose qualifications were otherwise similar but whose names signaled different gender and racial identities. In their randomized experiment, the models generally gave female candidates higher scores, while most gave black male candidates lower scores than comparable white male candidates. The authors emphasize that broad group averages can hide these opposing patterns: a model could favor female candidates and still disadvantage black men. They also found that the direction and size of score differences varied across social groups and models, and that results for Asian and Hispanic candidates differed across models. To estimate how scores might translate into hiring, the authors applied assumed score cutoffs. At an 80-point cutoff, GPT-3.5 Turbo’s scores corresponded to estimated hiring-probability differences for otherwise similar groups. These are estimates under the study’s rule, not observed hiring outcomes. The researchers used randomized resumes with characteristics drawn from real-world distributions, but tested English-language, entry-level applications in a US setting, using names to signal identity. The authors say further research is needed to identify the causes of the patterns and how they may change over time.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

## Original source

PubMed Central · Article

Original source date: Mar 12, 2025 (2025-03-12).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T13:28:28.417Z. First observed: 2026-11-16T22:56:25.147Z. Retrieved: 2026-11-17T13:27:32.155Z.

- [Measuring gender and racial biases in large language models: Intersectional evidence from automated resume evaluation](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Limits

The experiment used fictitious resumes and names to signal gender and racial identity; it does not assess identity information beyond those names. The authors caution that the findings are specific to an English-speaking US context and may not extend to other linguistic or cultural settings. The study covers entry-level positions and standardized resumes, so the authors say its results may not extend to high-skill or management hiring. The hiring-probability estimates depend on an assumed score cutoff; they are not observed hiring outcomes. The experiment was conducted at one snapshot in time, and the authors note that model behavior may evolve.


## Open question

Would the score differences persist when real applications are assessed with professional recruiters’ prompts?


## Source-stated

In a randomized resume-scoring experiment, the models generally assigned higher scores to female candidates and most assigned lower scores to black male candidates than to comparable white male candidates.

> the LLMs award higher assessment scores for female candidates with similar work experience, education, and skills, but lower scores for black male candidates with comparable qualifications

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source-stated

The authors found that aggregate comparisons can conceal different patterns at the intersection of gender and race.

> the higher scores for minorities are driven by those for female candidates (both black and white females), but the scores for black males are in fact significantly lower

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source-stated

At an 80-point cutoff, the authors estimate score-based hiring-probability differences for GPT-3.5 Turbo; these are estimates under an assumed rule, not observed hiring outcomes.

> with an 80-point cutoff, which implies an average hiring probability of ∼35%, black female applicants could expect a 1.7 percentage-point greater probability of being hired

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source-stated

The authors say their results may not extend beyond English-speaking US settings or entry-level hiring, and identity was inferred from names.

> our findings are relevant only to an English-speaking context, specifically within the United States, and may not extend to other linguistic or cultural settings

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source and contributor context

Jiafu An, Difang Huang, Chen Lin, Mingzhu Tai · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

Media declared by the original publisher; availability checked separately from the editorial reading.

- [Regression coefficients: GPT-3.5 Turbo score differences across social groups. This figure presents the regression coefficients that compare the score differen… · PubMed Central](https://cdn.ncbi.nlm.nih.gov/pmc/blobs/1cee/11937954/1b031d780666/pgaf089f1.jpg)

## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## PubMed Central

PubMed Central is a US National Library of Medicine archive providing full-text biomedical and life sciences research articles.

- [Explore this context](/entities/pubmed-central)
- [Reference](https://pmc.ncbi.nlm.nih.gov/)

## TIME

TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.

- [Explore this context](/entities/time)
- [Reference](https://time.com/)

## Editorial provenance

Version 1. Publication ID: c9669bb2-5895-4399-8b9c-61b443227576.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-17T13:28:28.417Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 6d9d1da4e5d32591a63c2640581b4eef96da6680f806cc5850a6220cd4b4094f. Source text hash: ca7ae3ffcf047d449d45d6b8bfc7d99454c99427ea6cedf21c427e1a2644102e.


## Revision and source-check history

### Version 1 — 2026-11-17T13:28:28.417Z

Jiafu An and colleagues report that five large language models gave different resume scores to fictitious applicants whose qualifications were otherwise similar but whose names signaled different gender and racial identities. In their randomized experiment, the models generally gave female candidates higher scores, while most gave black male candidates lower scores than comparable white male candidates. The authors emphasize that broad group averages can hide these opposing patterns: a model could favor female candidates and still disadvantage black men. They also found that the direction and size of score differences varied across social groups and models, and that results for Asian and Hispanic candidates differed across models. To estimate how scores might translate into hiring, the authors applied assumed score cutoffs. At an 80-point cutoff, GPT-3.5 Turbo’s scores corresponded to estimated hiring-probability differences for otherwise similar groups. These are estimates under the study’s rule, not observed hiring outcomes. The researchers used randomized resumes with characteristics drawn from real-world distributions, but tested English-language, entry-level applications in a US setting, using names to signal identity. The authors say further research is needed to identify the causes of the patterns and how they may change over time.

## Original source

PubMed Central · Article

Original source date: Mar 12, 2025 (2025-03-12).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T13:28:28.417Z. First observed: 2026-11-16T22:56:25.147Z. Retrieved: 2026-11-17T13:27:32.155Z.

- [Measuring gender and racial biases in large language models: Intersectional evidence from automated resume evaluation](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Limits

The experiment used fictitious resumes and names to signal gender and racial identity; it does not assess identity information beyond those names. The authors caution that the findings are specific to an English-speaking US context and may not extend to other linguistic or cultural settings. The study covers entry-level positions and standardized resumes, so the authors say its results may not extend to high-skill or management hiring. The hiring-probability estimates depend on an assumed score cutoff; they are not observed hiring outcomes. The experiment was conducted at one snapshot in time, and the authors note that model behavior may evolve.


## Open question

Would the score differences persist when real applications are assessed with professional recruiters’ prompts?


## Source-stated

In a randomized resume-scoring experiment, the models generally assigned higher scores to female candidates and most assigned lower scores to black male candidates than to comparable white male candidates.

> the LLMs award higher assessment scores for female candidates with similar work experience, education, and skills, but lower scores for black male candidates with comparable qualifications

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source-stated

The authors found that aggregate comparisons can conceal different patterns at the intersection of gender and race.

> the higher scores for minorities are driven by those for female candidates (both black and white females), but the scores for black males are in fact significantly lower

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source-stated

At an 80-point cutoff, the authors estimate score-based hiring-probability differences for GPT-3.5 Turbo; these are estimates under an assumed rule, not observed hiring outcomes.

> with an 80-point cutoff, which implies an average hiring probability of ∼35%, black female applicants could expect a 1.7 percentage-point greater probability of being hired

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source-stated

The authors say their results may not extend beyond English-speaking US settings or entry-level hiring, and identity was inferred from names.

> our findings are relevant only to an English-speaking context, specifically within the United States, and may not extend to other linguistic or cultural settings

- [Supporting source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11937954/)

## Source and contributor context

Jiafu An, Difang Huang, Chen Lin, Mingzhu Tai · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

Media declared by the original publisher; availability checked separately from the editorial reading.

- [Regression coefficients: GPT-3.5 Turbo score differences across social groups. This figure presents the regression coefficients that compare the score differen… · PubMed Central](https://cdn.ncbi.nlm.nih.gov/pmc/blobs/1cee/11937954/1b031d780666/pgaf089f1.jpg)

## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## PubMed Central

PubMed Central is a US National Library of Medicine archive providing full-text biomedical and life sciences research articles.

- [Explore this context](/entities/pubmed-central)
- [Reference](https://pmc.ncbi.nlm.nih.gov/)

## TIME

TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.

- [Explore this context](/entities/time)
- [Reference](https://time.com/)

## Editorial provenance

Version 1. Publication ID: c9669bb2-5895-4399-8b9c-61b443227576.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-17T13:28:28.417Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: 6d9d1da4e5d32591a63c2640581b4eef96da6680f806cc5850a6220cd4b4094f. Source text hash: ca7ae3ffcf047d449d45d6b8bfc7d99454c99427ea6cedf21c427e1a2644102e.


### uncertain — 2026-11-17T13:28:28.417Z — anthropic/claude-opus-5.5

Publication: c9669bb2-5895-4399-8b9c-61b443227576

attribution — pass: The reading attributes the study to Jiafu An and colleagues, accessed via PubMed Central (PNAS Nexus article). Findings, cautions and estimates are framed as the authors' claims, not as established facts.

> Jiafu An
1, Difang Huang
2,2,✉, Chen Lin
3,✉, Mingzhu Tai

summary — pass: The title, summary and question faithfully convey the randomized design, the pro-female and anti-black-male pattern, the cross-model variation, the cutoff-based estimates and the scope limits. A general reader can follow it. The question is fair: the authors used brief baseline prompts and fictitious resumes. Optional repairs: state the roughly 1–3 percentage-point magnitude, and note the authors' conjecture that the debiasing stage likely drives the female advantage. Neither omission distorts the reading.

> These biases may result in ∼1–3 percentage-point differences in hiring probabilities for otherwise similar candidates at a certain threshold

date — uncertain: The displayed date, 2025-03-12, matches the PNAS Nexus publication. However, the source says the manuscript was first posted as an arXiv preprint. The ID prefix 2403 suggests March 2024, but the printed ID appears truncated, so the exact first-posting date cannot be confirmed. The displayed date may therefore reflect a later journal version rather than the work's first public release.

> This manuscript was posted on a preprint: http://arxiv.org/abs/2403.1528.

support — pass: Each claim's excerpt is verbatim and supports it. The reading correctly describes the hiring-probability figures as estimates under an assumed cutoff rule, not observed hiring outcomes. The listed limitations (names as identity signals, US English-speaking context, entry-level jobs, a single point in time) all match the authors' stated caveats. No unsupported benefits, harms or forecasts are added.

> Lastly, our experiment is only conducted at one snapshot of time. As AIs continue to evolve and learn, the way they consider social identities in decision-making can also evolve.

media — pass: The selected media is the verified direct link to the original article on PubMed Central. This is the most direct access to the source.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/0383417639bc229d88a95329385189a38ca2b7a0b30e143744bf9a2873f0d03f/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/pubmed-central-an-and-colleagues-find-intersecting-biases-in-03834176.json)
