Model Explanations for Gender and Ethnicity Bias Mitigation in AI-Generated Narratives

Large Language Models (LLMs) are increasingly utilized in diverse applications, ranging from professional content creation to decision-making systems. However, their outputs often amplify the biases present in their training data, perpetuating stereotypes and reinforcing societal inequities, particularly regarding gender and ethnicity. Such biases can cause tangible harm, especially for underrepresented groups, and require awareness and effective mitigation strategies. This work explores gender and ethnicity representation in narratives created by generative AI describing 25 occupational fields defined by the U.S. Bureau of Labor Statistics. We examine three large language models (LLMs)--Llama 3.1 70B Instruct, Claude 3.5 Sonnet, and GPT 4.0 Turbo. Employing a novel approach that leverages model-generated explanations, we assess bias before and after mitigation using two metrics: Demographic Parity Ratio (DPR) and Total Variation Distance (TVD). Our findings reveal that incorporating model explanations significantly improves demographic representation, reducing biases by 2%-20% across different models and occupations. Qualitative analysis of the generated stories indicates high levels of creativity, coherence, and inclusivity, demonstrating the potential of targeted interventions to produce equitable narratives. This research contributes a robust dataset of occupational narratives and a systematic framework for bias mitigation, advancing the understanding of LLM behavior and promoting ethical AI development. By aligning explainability with equity, this work underscores the critical role of transparency and accountability in the deployment of generative AI systems.