Can LLMs Hire Fairly? Racial Bias in Resume Screening

A new audit of fourteen large language models finds that the direction of racial bias in automated resume screening has reversed across model generations, with a 2023-vintage model favoring White candidates while every model released in 2024 or later shows either no gap or a significant preference for Black candidates, according to a preprint posted to arXiv on June 27, 2026 [1]. The study applied the paired-resume methodology developed by Kline, Rose, and Walters to 24,024 paired postings across the fourteen models [1][2]. The sole 2023-vintage model, GPT-3.5-turbo, reproduced a pro-White callback gap of +2.12 percentage points, significant at the 1% level [1][3]. Every model released in 2024 or after exhibited either a null gap or a statistically significant pro-Black reversal, with gaps ranging from −0.37 to −3.01 percentage points [3][4]. The same pattern held on the gender axis: GPT-3.5-turbo favored male candidates by +1.92 percentage points, while all 2024-and-later models uniformly favored female candidates or showed no gap [3][4]. The reversal was not confined to a single provider or model family. It appeared across models from OpenAI, Anthropic, Meta, Google, xAI, DeepSeek, Alibaba, and Zhipu, and proved robust to posting-level cluster-bootstrap inference [3][4]. Within the OpenAI lineage alone, the race gap moved from +2.12 percentage points in 2023 to −0.61 percentage points in 2024, and then to null in 2026 [3][4]. The authors hypothesize that successive rounds of post-training alignment shifted model behavior from reproducing pro-White patterns found in pretraining data to actively favoring minority-coded names, and more recently toward neutrality as alignment techniques matured [3][4]. They caution that neither the original pro-White bias nor its pro-Black reversal satisfies the requirement of equal treatment regardless of race or gender [3][4]. The findings land amid broader scrutiny of AI hiring tools. A separate 2024 audit of Massive Text Embedding models found that White-associated names were significantly favored in 85.1% of cases, while female-associated names were favored in only 11.1% of cases, with Black males disadvantaged in up to 100% of cases [5][6]. Another recent study introduced the term "Illusion of Neutrality" to describe models that appear unbiased only because they lack the competence to make meaningful evaluations, relying instead on superficial keyword matching [8].