💬 NLP

Can LLMs Hire Fairly? Racial Bias in Resume Screening

This paper audits fourteen large language models and finds that while a 2023 model exhibited pro-White hiring bias similar to human discrimination, all models released in 2024 or later demonstrated either no bias or a significant pro-Black reversal, marking a generational shift in algorithmic hiring fairness.

Original authors: Zhenyu Gao, Wenxi Jiang, Yutong Yan

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhenyu Gao, Wenxi Jiang, Yutong Yan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring for a job, but instead of a human looking at the resumes, you ask a super-smart computer robot to make the decision. You want to know: Is this robot fair?

This paper is like a massive "mystery shopper" experiment. The researchers hired 14 different AI robots (Large Language Models) to act as hiring managers. They gave each robot thousands of pairs of fake resumes that were identical in every way—same education, same work history, same age, same skills. The only thing that changed was the name on the resume.

Here is the story of what they found, told through a few simple analogies:

1. The "Old Robot" vs. The "New Robots"

Think of the AI models like different generations of smartphones.

The 2023 Model (The "Old Phone"): The researchers tested one older model (GPT-3.5-turbo). This robot acted exactly like the biased humans found in past studies. When it saw a name that sounded Black, it was less likely to say "Yes, call this person in" compared to a name that sounded White. It showed a clear preference for White names.
The 2024+ Models (The "New Phones"): Then, they tested 13 newer models released in 2024 and 2025. These robots did something surprising. They didn't just become neutral; they swung the other way. They started favoring Black names over White names.

The Analogy: Imagine a scale.

The old robot tipped the scale heavily to the White side.
The new robots tipped the scale heavily to the Black side.
The newest robots (from 2026) are trying to balance the scale perfectly, showing almost no preference either way.

2. The Same Story for Men and Women

They ran the exact same test but changed the names to see if the robots were biased against men or women.

The Old Robot: Preferred men over women.
The New Robots: Preferred women over men.
The Newest Robots: Are becoming neutral again.

3. Why Did This Happen?

The paper suggests this isn't a glitch; it's a result of how the robots were "taught" to behave.

The Training Data: The old robot learned from the internet as it was, which contains a lot of historical human bias (favoring White men).
The "Alignment" Fix: The companies that built the newer robots realized this was unfair. They added a special "training class" (called alignment) to teach the robots to be fair.
The Over-Correction: It seems the newer robots got a little too excited about being fair. Instead of just being neutral, they started actively favoring the groups that were previously discriminated against. It's like a teacher who, after seeing a student get bullied, decides to give that student extra credit on every test, even when they didn't earn it.

4. The Big Takeaway

The main lesson from this paper is that AI bias is not a fixed thing. It changes depending on which version of the robot you use and when it was built.

2023 Robots: Repeated the old human mistakes (favoring White men).
2024-2025 Robots: Tried to fix it but swung too far the other way (favoring Black women).
2026 Robots: Are getting closer to being truly neutral.

The Bottom Line:
The paper concludes that just because a tool is "AI" doesn't mean it's fair. Sometimes it's too biased one way, and sometimes it's too biased the other. The only way to know if a specific AI hiring tool is fair is to test it constantly, because the "personality" of the AI changes with every new update.

What the paper does NOT say:

It does not say these robots are currently being used by all companies (though it mentions a lawsuit against one company, Workday).
It does not say we should stop using AI.
It does not say the "pro-Black" bias is a good thing; the authors argue that any bias (favoring one group over another) is bad for fairness. The goal is a neutral scale, not a tipped one.

Technical Summary: Can LLMs Hire Fairly? Racial Bias in Resume Screening

Problem Statement

As Large Language Models (LLMs) become increasingly embedded in consequential economic decisions, specifically hiring, there is a critical need to determine whether these systems perpetuate existing patterns of human discrimination or reshape them. While legal challenges (e.g., the 2026 Workday class-action suit) have highlighted the potential for algorithmic discrimination, the empirical evidence regarding the specific nature and direction of bias in LLM-based hiring screeners remains limited. The central question addressed is whether LLMs, when acting as HR managers, exhibit racial and gender bias, and how this bias evolves across different model generations and providers.

Methodology

The study employs a large-scale paired-resume audit methodology, directly adapting the correspondence experiment design established by Bertrand and Mullainathan (2004) and scaled by Kline, Rose, and Walters (2022).

Experimental Design: The researchers constructed candidate profiles that were identical in all substantive attributes (education, employment history, age, gender) except for the candidate's first and last name. Names were drawn from established lists of distinctively Black and White names (Bertrand and Mullainathan, 2004; Kline et al., 2022).
Data Source: The audit utilized 6,007 real US entry-level job postings from the Revelio Labs universe, sampled to match the Fortune-500 firm distribution.
Model Panel: Fourteen mainstream LLMs were tested, spanning release dates from May 2023 to April 2026. The panel includes models from eight providers (OpenAI, Anthropic, Meta, Google, xAI, DeepSeek, Alibaba, and Zhipu) with parameter sizes ranging from 8B to 397B.
Prompting and Scoring: Each model was prompted with a standardized system instruction to act as an HR manager and respond with a single word ("yes" or "no") indicating whether to advance a candidate. Temperature was set to zero to ensure deterministic outputs.
Scale: The study generated 24,024 paired postings per model for the race axis and 48,048 paired postings per model for the gender axis.
Statistical Analysis: The primary metric was the callback gap ( $\Delta$ ), calculated as the difference in "yes" rates between advantaged and disadvantaged groups. Significance was assessed using McNemar's test on discordant pairs, with cluster-robust inference (bootstrap and standard errors) applied at the posting level to account for within-posting dependence.

Key Contributions

Systematic Multi-Model Audit: This work provides the first systematic audit of LLM hiring discrimination at a scale comparable to the field experiments that established human baselines, covering 14 models on the race axis and 13 on the gender axis.
Documentation of Sign Reversal: The study documents a qualitative shift in the direction of algorithmic discrimination across model generations, moving from pro-White/pro-Male bias in 2023 models to pro-Black/pro-Female bias (or null gaps) in 2024+ models.
Correlation of Bias Axes: The paper demonstrates that race and gender bias are positively correlated across models; models favoring Black-coded names also tend to favor female-coded names.

Key Results

Race Discrimination

2023 Vintage (GPT-3.5-turbo): The sole 2023 model reproduced the pro-White callback gap found in human field experiments. White-coded names received callbacks 2.12 percentage points (pp) more often than Black-coded names (significant at the 1% level). This magnitude exceeded the 1.6 pp gap reported by Kline et al. (2022).
2024+ Vintage: Every model released in 2024 or later exhibited either a null gap or a statistically significant pro-Black reversal. Significant pro-Black gaps ranged from -0.37 pp to -3.01 pp (e.g., Llama-3.1-8B-Instruct showed -3.01 pp). No 2024+ model showed a significant pro-White gap.
Trajectory: Within model families, the trend was consistent. OpenAI's lineage moved from +2.12 pp (pro-White) in GPT-3.5-turbo to -0.61 pp (pro-Black) in GPT-4o-mini, and finally to a null gap in GPT-5.4-mini.

Gender Discrimination

2023 Vintage: GPT-3.5-turbo favored male candidates by 1.92 pp (significant at the 1% level).
2024+ Vintage: Ten of the twelve remaining models showed a significant pro-Female gap, ranging from -0.18 pp to -5.80 pp. Two models (Gemini-2.5-flash and GPT-5.4-mini) showed null gaps.
Correlation: The race and gender axes were positively correlated. The sole pro-White model (GPT-3.5-turbo) was also the sole pro-Male model. Conversely, models with the strongest pro-Black bias (e.g., Claude 3 Haiku) also exhibited the strongest pro-Female bias.

Robustness

The findings remained robust under cluster-robust inference and Bonferroni corrections for multiple comparisons. While a few borderline results (e.g., Gemma-4-31B gender gap) lost significance under strict correction, the core finding of a sign reversal across generations held firm.

Significance and Claims

The paper claims that the direction of algorithmic hiring bias is not a fixed property of language models but varies systematically with model vintage, provider, and scale.

Evolution of Alignment: The observed trajectory—from reproducing pretraining data biases (pro-White) to actively favoring minority-coded names (pro-Black), and potentially toward neutrality in the latest models—is attributed to changes in post-training alignment procedures between 2023 and 2024.
Fairness Implications: The authors argue that neither the original pro-White bias nor the subsequent pro-Black reversal is desirable from a fairness standpoint. A hiring screener that systematically favors one group over another, regardless of direction, fails the basic requirement of equal treatment.
Regulatory Relevance: The results provide empirical evidence for regulators (such as those enforcing the EU AI Act or NYC Local Law 144) that the magnitude and direction of bias depend on specific deployment choices (model version, training date, prompting). The authors conclude that systematic auditing of the kind conducted here is essential for regulatory compliance and public accountability as AI tools become ubiquitous in employment screening.