The rapid advancement of Generative Artificial Intelligence (AI), particularly Large Language Models (LLMs), has transformed the landscape of automated text generation across domains such as recruitment, education, media, and decision-support systems. Despite their remarkable capabilities, these models may inadvertently encode, reproduce, and amplify existing societal biases present in training data. This study explores algorithmic bias within language generation frameworks, with a particular emphasis on identifying and quantifying gender-related disparities in AI-generated content. The research proposes a systematic bias evaluation framework grounded in statistical and probabilistic methods. Key metrics include mean bias, mean absolute bias, sentiment distribution analysis, and divergence-based measures such as Kullback–Leibler divergence to assess distributional imbalances across demographic attributes. By analyzing large-scale AI-generated textual datasets, the study aims to detect measurable patterns of bias and evaluate how prompt construction, model architecture, and training data influence output disparities. Furthermore, the work examines cross-model behavior among leading proprietary and open-source LLMs and integrates interpretable embedding techniques to enhance transparency in bias detection and mitigation. The expected contribution of this research lies in developing a mathematically rigorous bias quantification pipeline and offering practical strategies for fairness-aware language generation. Ultimately, the study seeks to provide a scalable framework for evaluating and reducing algorithmic bias in generative AI systems, contributing to more equitable and responsible AI deployment.
Sharavaiah et al. (2026) studied this question.