Large language models show gender sensitivity in candidate evaluations
LLMs Cannot Be Neutral: Gender Sensitivity Across Candidates, Evaluators, and Models
Computation and LanguageHuman-Computer Interaction
Summary
Large language models are sometimes used to judge candidates for jobs, but they may treat people differently based on gender. The researchers tested several models by having them evaluate fictional profiles with different gender labels. They found that specifying a candidate’s gender slightly changed the likelihood of a positive decision, and models varied in how sensitive they were. The study also showed the importance of looking at individual decisions, not just overall trends, to understand bias.
What this means in practice
- •For hr technology teams: Improve fairness checks by assessing gender sensitivity in AI-based candidate screening tools.
- •For ai ethics officers: Analyze how gender labels impact evaluation outcomes to guide responsible AI use in professional hiring.
Authors
Danial Amin, Joni Salminen, Bernard J. Jansen
Abstract
Large language models (LLMs) are increasingly used to evaluate candidates for professional roles, raising concerns about their sensitivity to demographic identities. We examine how candidate and evaluator gender assignments affect decisions by LLM-as-a-Judge (LaJ) models and decision language models (DLMs). In a hypothetical city council voting task, eight models evaluated 240 synthetic professional narratives from NVIDIA's US and French Nemotron-Personas datasets under independently varied gender assignments, producing 30,720 judgments. Explicit candidate gender labels increased Yes-vote probability by 1.05-1.73 percentage points relative to unspecified gender, while only the nonbinary evaluator assignment showed a significant increase (1.26 points). Identity matching showed no significant overall association. Yes-vote rates varied across models (81.3-97.2%), and 2.5-3.3% of decisions reversed relative to the unspecified-identity baseline. Under direct substitutions between explicit gender identities, DLMs exhibited fewer decision reversals than LaJ models (1.18% vs. 3.38%), indicating lower observed decision-level gender sensitivity among the tested DLMs. These findings demonstrate that gender sensitivity depends on candidate and evaluator identities and the evaluation model, while highlighting the importance of examining individual decision consistency alongside aggregate differences.