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

Research from arXiv highlights that the widespread use of large language models (LLMs) in hiring introduces a risk of 'monocultural biases,' which could lead to greater systemic exclusion for specific demographic groups. The study indicates that post-trained LLMs, compared to their base versions, are significantly more likely to negatively impact certain applicant groups, with a notable reduction in callbacks for older applicants across multiple models.

Why it matters

The findings underscore a critical risk associated with the adoption of AI in human resources, specifically the potential for uniform and systemic discriminatory outcomes if not properly managed. This uniformity of bias can entrench existing inequalities, affecting labor market diversity and organizational access to a broad talent pool.

Key insights

• Employers are increasingly adopting large language models (LLMs) for automated hiring processes.
• The research investigates 'monocultural biases,' where widespread LLM deployment homogenizes biases across the labor market, increasing systemic exclusion for certain demographics.
• Post-trained LLMs are 3.6% less likely to provide callbacks to older applicants compared to their base models.
• This negative shift regarding older applicants was observed in eight out of ten evaluated models.
• Post-trained models demonstrate a higher correlation in their hiring decisions than base models, suggesting a homogenization of biases.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.22169