Systematic literature review on bias mitigation in generative AI

Abstract

In the era of rapid technological advancement, Artificial Intelligence (AI) is a transformative force, permeating diverse facets of society. However, bias concerns have gained prominence as AI systems become integral to decision-making processes. Bias can exert significant and extensive consequences, influencing individuals, groups, and society. The presence of bias in generative AI or machine learning systems can produce content that exhibits discriminating tendencies, perpetuates stereotypes, and contributes to inequalities. Artificial intelligence (AI) systems have the potential to be employed in various contexts that involve sensitive settings, where they are tasked with making significant judgements that can have profound impacts on individuals' lives. Consequently, it is important to establish measures that prevent these decisions from exhibiting discriminating tendencies against specific groups or populations. This exclusive exploration embarks on a comprehensive journey through the nuanced landscape of bias in AI, unravelling its intricate layers to discern different types, pinpoint underlying causes, and illuminate innovative mitigation strategies. Delving deeper, we investigate the roots of bias in AI, revealing a complex interplay of historical legacies, societal imbalances, and algorithmic intricacies. Unravelling the causes involves exploring unintentional reinforcement of existing biases, reliance on incomplete or biased training data, and the potential amplification of disparities when AI systems are deployed in diverse real-world scenarios. Various domains such as text, image, audio, video and more significant advancements in Generative Artificial Intelligence (GAI) were evidenced. Multiple challenges and proliferation of biases occur in different perspectives considered in the study. Against this backdrop, the exploration transitions to a proactive stance, offering a glimpse into cutting-edge mitigation strategies. Diverse and inclusive datasets emerge as a cornerstone, ensuring representative input for AI models. Ethical considerations throughout the development lifecycle and ongoing monitoring mechanisms prove pivotal in mitigating biases that may arise during training or deployment. Technical and non-technical strategies come to the forefront of pursuing fairness and equity in AI. The paper underscores the importance of interdisciplinary collaboration, emphasising that a collective effort spanning developers, ethicists, policymakers, and end-users is paramount for effective bias mitigation. As AI continues its ascent into various spheres of our lives, understanding, acknowledging, and addressing bias becomes an imperative. This exploration seeks to contribute to the discourse, fostering a deeper comprehension of the challenges posed by bias in AI and inspiring a collective commitment to building equitable, trustworthy AI systems for the future.