Bias Detection and Mitigation in Generative AI Systems

Abstract

Generative AI systems -- large language models, image generators, and multimodal models -- encode and amplify societal biases present in their training data, producing outputs that systematically disadvantage or misrepresent individuals on the basis of protected characteristics including gender, race, ethnicity, religion, age, and disability. Unlike bias in classification-based AI where individual decisions can be audited, generative AI bias manifests across open-ended output spaces that are difficult to exhaustively test and that interact with user prompts in complex ways to produce varying bias profiles. This paper proposes the Generative AI Bias Evaluation and Mitigation (GABEM) framework, a systematic methodology for detecting, quantifying, and mitigating bias in generative AI systems across five bias dimensions: representational bias, stereotyping bias, sentiment bias, opportunity bias, and intersectional bias. GABEM integrates three bias detection methods -- distributional analysis (DA), counterfactual probing (CP), and generative bias audit (GBA) -- with three mitigation strategies -- counterfactual data augmentation (CDA), bias-aware reinforcement learning from human feedback (BA-RLHF), and debiased decoding constraints (DDC). GABEM is evaluated on three generative AI systems: a text generation LLM (LLaMA-2-7B), a text-to-image diffusion model (Stable Diffusion XL), and a multimodal generation model (LLaVA-1.5-13B), using a comprehensive bias evaluation suite of 12 benchmarks across the five bias dimensions. Results demonstrate that the GABEM mitigation suite reduces mean bias scores by 48.4% (SD = 6.2%) while maintaining 94.6% of baseline generation quality. Intersectional bias -- the compounded bias affecting individuals at multiple protected characteristic intersections -- shows the largest mitigation benefit (62.8% reduction) and was previously unaddressed in generative AI bias literature. The study contributes the GABEM framework, a Generative Bias Score (GBS) composite metric, and the largest cross-system generative AI bias evaluation conducted to date.