The rapid proliferation of generative artificial intelligence, encompassing large language models (LLMs) such as GPT-4 and Gemini as well as text-to-image generators including DALL·E and Stable Diffusion, has introduced unprecedented capabilities alongside profound ethical hazards. This paper investigates how societal biases encoded in training corpora are systematically perpetuated, amplified, and reified by these systems. Drawing on three domain case analyses, namely automated hiring screening, algorithmic credit scoring, and predictive law-enforcement tools, we synthesise documented evidence of disparate impact across race, gender, and socioeconomic strata, and foreground intersectional harm as a distinct and frequently under-measured dimension of that impact. We are explicit that these analyses are narrative syntheses of secondary sources rather than original longitudinal data collection. We then propose a tripartite governance framework comprising (1) pre-ingestion dataset auditing, (2) in-training fairness constraints via adversarial debiasing, and (3) post-deployment algorithmic auditing protocols informed by Explainable AI (XAI) techniques. We situate these technical proposals within an explicit ethical analysis, applying Rawlsian fairness principles and Sen’s capability approach to specific cases, and situating these within the political economy of AI governance. Our central argument is that no single debiasing technique is sufficient, that several techniques are themselves in tension with privacy and with one another, and that meaningful fairness demands coordinated intervention across the model lifecycle and the wider sociotechnical system. We conclude with policy recommendations for regulators, model developers, and civil society stakeholders, and with reflection on how these dynamics manifest beyond the Global North.