Bias in recruitment systems utilizing large language models

AI-powered recruitment systems are becoming increasingly common, offering greater efficiency in hiring processes. However, these systems may unintentionally perpetuate biases, leading to unfair hiring decisions. This study investigates various types of biases such as gender, racial, and age in recruitment systems using several large language models (LLMs) including BERT, GPT-2, and GPT-Neo. Bias in AI-driven recruitment models is assessed using the Word Embedding Association Test (WEAT) metric. The experimental results reveal varying levels of bias across LLMs and significant biases in gender and racial associations, highlighting potential disparities in AI-driven hiring recommendations. To this end, this paper underscores the need for bias mitigation strategies in AI-based recruitment tools and advocates for transparent and equitable hiring practices. The source code is publicly available at https://github.com/soujanyaachandra/Recruitment_Bias_Analysis.