Assessing Bias and Fairness in AI Language Models: A Systematic Review of Current Approaches

Authors

  • Aisha AKANBI Airforce Institute of Technology, Kaduna, Nigeria
  • Fatima Binta Abdullahi University of Abuja
  • Muhammed Bello Akanbi University of Abuja
  • Adamu Yusuf Atumosha University of Abuja

Keywords:

Algorithmic Fairness, Algorithmic Bias Mitigation, Bias Detection, Debiasing, Demographic Parity, Equalized Odds

Abstract

Artificial intelligence language models have demonstrated remarkable capabilities across diverse natural language processing tasks; however, systematic biases embedded in training corpora continue to propagate discriminatory outcomes across demographic groups. As large language models (LLMs) are increasingly deployed in high-stakes domains including healthcare, criminal justice, hiring, and financial services, their potential to amplify societal inequities has attracted urgent scholarly and regulatory attention. Despite a rapidly growing body of empirical work, no prior study has comprehensively synthesized bias research across all major LLM families using rigorous systematic review methodology. This paper addresses that gap by presenting a comprehensive systematic review of 147 peer-reviewed publications (2018–2024) examining bias detection, quantification, and mitigation strategies in LLMs, conducted in strict accordance with PRISMA 2020 guidelines. We synthesized empirical findings across gender, racial, socioeconomic, and intersectional bias dimensions, and introduce a novel five-tier Bias Severity Classification Framework (BSCF) that standardizes how bias magnitude, demographic scope, and harm potential are categorized and reported. The BSCF tiers range from Negligible (T1, demographic score delta ≤ 0.02) to Critical (T5, delta > 0.30 with intersectional impact), providing a model-agnostic vocabulary for both research reporting and regulatory compliance. Quantitative meta-analysis using a random-effects model across 89 debiasing interventions reveals that state-of-the-art methods reduce measured bias by 23–67% while incurring 2–11% performance degradation on downstream tasks, with heterogeneity index I² = 71.4%, indicating substantial variability across intervention families. Our analysis of benchmark datasets including WinoBias, StereoSet, BBQ, CrowS-Pairs, and BOLD further reveals that no single benchmark captures the full bias landscape, underscoring the need for multi-benchmark evaluation suites. We identify three critical research gaps: multilingual bias coverage, temporal drift of bias metrics, and ecological validity of laboratory evaluations. We further present a standardized evaluation pipeline and actionable recommendations for equitable AI deployment, with the BSCF proposed as a reporting standard for future empirical work.

Downloads

Published

2026-06-24