Bias and Fairness in AI Algorithms: A Systematic Literature Review

Abstract

AI algorithms are more deeply integrated into high-stakes decision systems within healthcare, criminal justice, finance, and hiring. Although efficient, these systems often replicate and fuel societal biases that yield outcomes skewed against disadvantaged groups within the population. This systematic literature review (SLR) aims, in accordance with the PRISMA 2020 guidelines, to synthesize peer-reviewed evidence regarding the nature of bias as well as its measurement and mitigation in AI algorithms. We identified 936 papers through a systematic search of Google Scholar and PubMed (2015–2025), which retrieved all records of interest. Two levels of screening resulted in the inclusion of 121 studies in this review. The analysis identified three main types of bias mitigation: in-processing, pre-processing, and post-processing approaches. The most commonly used fairness metrics are demographic parity, equalized odds, and equal opportunity. Examples of application domains in which you are well represented include healthcare, lending, criminal justice, and recruitment. Across studies, we also observed a reproducible accuracy-fairness trade-off; for every 1–5% improvement in model fairness, at most an equal accuracy drop was typically noted. Such a failure highlights important limitations in intersectional fairness studies, minimal validation of practical implementation, and lack of a universal metric approach to ensure approximate levels of equity. Future studies should include domain-specific fairness frameworks, intersectional methodological designs, and longitudinal evaluations of deployed AI systems.