Monocultural Biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring

As employers increasingly use large language models to automate hiring, this paper examines the risk of "monocultural biases": when widespread deployment of LLMs homogenizes biases across the labor market, certain demographic groups face greater systemic exclusion.

The authors measure hiring biases across the base and post-trained versions of ten LLMs, in order to identify whether pre-training or post-training is the stage that produces monocultural biases.

Compared with their base models, post-trained models are 3.6% less likely to call back older applicants, a negative shift seen in eight of the ten models evaluated. Post-trained models also make far more correlated decisions than base models, likely driven by human capital traits such as skills or college major. That greater consensus among models raises global systemic exclusion rates from 5.6% to 17.3% and worsens demographic inequalities, with intersectional systemic exclusion rates ranging from 12.2% to 21.7% for post-trained models. The inequality is primarily driven by age-based discrimination that post-training exacerbates.

Overall, the results indicate that post-training techniques may improve models' ability to select the best applicants, but can also raise systemic inequality risks for those at the margin by uniformly introducing new biases.