Authors: Ms. Kajal Yadav, Kavita Ahirwar

Abstract: This research paper focuses on the mechanics, implications, and demographic effects of bias in contemporary human resources recruitment processes. As more companies integrate AI technology to improve the efficiency of recruiting large quantities of applicants, automated technologies, such as Natural Language Processing (NLP)-based ATS and Computer Vision/Audio-based AVI, have changed hiring processes significantly. Though the vendors of these tools claim them to be unbiased tools for overcoming human cognitive biases in evaluation, numerous empirical studies show that these technologies tend to perpetuate and exacerbate the historical employment biases. Through the use of an empirical research design, the study analyses applicant evaluation outputs (N=450 recruitment profiles interactions in Fortune 500 hiring processes) using traditional metrics of algorithmic fairness, which include Disparate Impact Ratio (DIR) based on EEOC's Four-Fifths rule, Demographic Parity, and Equalized Odds. The study uses systematic analysis of two main AI hiring pipelines: semantic resume screeners based on BERT and LLM embeddings, as well as affective/vocal AI video interview analyzers. The results show that the use of NLP-based resume parsers displays severe forms of proxy discrimination by punishing applicants belonging to minority demographic groups through implicit semantic associations with geographical zip code, university tier classification, and employment gap expressions. Moreover, facial expression analysis and vocal intonation detection during video interviews have shown substantial variance in systematic errors, which disadvantage non-native speakers, individuals with neurodevelopmental conditions, and racial minorities owing to the training data set distribution norms. Quantitative assessment indicates that the use of an unchecked AI screening system produces Disparate Impact Ratio below 0.72 for protected groups, which is inadequate according to the legal regulations. On the other hand, using pre-processing techniques, adversarial training of subnetworks, and Human-in-the-loop auditing results in Disparate Impact Ratio of 0.88.

DOI: http://doi.org/10.5281/zenodo.21720217