Academic research on AI credit scoring delivers strong predictive results but treats fairness and explainability as separate issues, leaving scarce guidance for regulated, real‑world use; the literature rarely evaluates integrated approaches that meet oversight and accountability requirements.

Main Finding

A systematic literature review of 43 peer‑reviewed studies (2020–2025) finds that AI models for credit scoring deliver strong predictive gains—notably for thin‑file applicants—but research and practice largely treat predictive performance, fairness, and explainability as separate objectives. There is limited evidence on methods and governance that jointly optimize these dimensions for deployment in regulated, high‑stakes lending contexts.

Key Points

• Scope and evidence base
  • 43 peer‑reviewed journal and conference articles (published 2020–2025), retrieved from IEEE Xplore, Scopus, Web of Science, and ScienceDirect (last searched 30 September 2025).
  • Inclusion required treatment of at least one of: performance, fairness, explainability in AI‑based credit scoring.
• Core empirical finding
  • AI approaches improve predictive accuracy, especially for thin‑file or sparse‑data borrowers, aiding credit access and risk assessment.
• Governance and adoption limits
  • Concerns over fairness, lack of transparent explanations, and regulatory accountability are major barriers to deployment in high‑stakes decisions.
• Fragmentation in the literature
  • Most studies address performance, fairness, or explainability in isolation; few evaluate their trade‑offs or integrate solutions in regulated deployment scenarios under human oversight.
• Quality appraisal
  • Review guided by the Relevance, Rigor, Reproducibility, and Quality (3Rs&Q) framework; both empirical and conceptual contributions were synthesized.
• Practical takeaway
  • The literature offers actionable recommendations (technical and governance) but empirical validation of integrated approaches in real‑world, regulated settings remains sparse.

Data & Methods

• Review type: Systematic literature review (SLR).
• Search and selection
  • Databases: IEEE Xplore, Scopus, Web of Science, ScienceDirect.
  • Time window: 2020–2025; last search conducted 30 September 2025.
  • Article types: Peer‑reviewed journal and conference papers.
  • Inclusion criterion: AI‑driven credit scoring applied to consumer/lending contexts addressing performance, fairness, or explainability.
• Appraisal framework: Relevance, Rigor, Reproducibility, and Quality (3Rs&Q) used to evaluate study contributions and map gaps.
• Synthesis approach: Comparative analysis across studies to identify common methods, metrics, regulatory considerations, and evidence gaps; distilled both empirical results and conceptual recommendations.

Implications for AI Economics

• Market efficiency and credit access
  • Improved predictive performance, especially for thin‑file borrowers, can increase credit supply and reduce information frictions, potentially expanding financial inclusion.
  • However, unaddressed fairness issues risk unequal access and biased pricing, counteracting inclusion gains.
• Distributional and welfare effects
  • If fairness and explainability are not addressed, AI deployment may exacerbate existing disparities (e.g., by protected attribute proxies), with adverse welfare and reputational costs for lenders.
• Regulatory and compliance costs
  • Requirements for explainability, auditability, and non‑discrimination create compliance burdens that affect the cost‑benefit calculus of AI adoption—smaller lenders may be disproportionately impacted.
• Competition and innovation
  • Firms that successfully integrate multi‑objective solutions (performance + fairness + explainability) can gain competitive advantage; fragmentation in research slows diffusion of such integrative practices.
• Systemic risk and model governance
  • Widespread use of similar black‑box models without robust oversight increases correlated model risk; regulators and market participants should prioritize monitoring and stress testing.
• Research and policy priorities
  • Need for economic research quantifying trade‑offs (e.g., accuracy vs. fairness vs. mortgage pricing), field experiments on deployment outcomes, and cost‑benefit analyses of regulatory instruments (e.g., disclosure mandates, audit regimes).
• Practical recommendations for stakeholders
  • Adopt multi‑objective model design (jointly optimizing accuracy, fairness, explainability).
  • Use transparent documentation (model cards, datasheets), rigorous subgroup evaluation, and post‑deployment monitoring.
  • Implement human‑in‑the‑loop processes, audit trails, and regulatory alignment to reduce adoption barriers and mitigate distributional harms.

If you want, I can: (a) extract example metrics and algorithms used across the reviewed studies, (b) draft a short checklist for lenders to operationalize the actionable guidance, or (c) outline a research agenda for economists studying the joint trade‑offs. Which would you prefer?