Bridging AI and Ethics: An LLM-Based Framework for Transparent and Inclusive Credit Decisions

Abstract

1) Research Context: The integration of advanced Machine Learning (ML) models into Intelligent Information Systems (IS) has created highly accurate but opaque "black-box" systems, especially in sensitive domains like credit scoring. 2) Scientific and/or Practical Problem: This opacity undermines user trust, can perpetuate algorithmic bias, and challenges regulatory compliance (e.g., LGPD, GDPR). This creates a critical gap between AI’s technical power and the socio-technical need for accountability in IS. 3) Proposed Solution and/or Analysis: We propose and validate a two-layer framework that uses Large Language Models (LLMs) to translate technical outputs from Explainable AI (XAI) methods, like SHAP and LIME, into actionable, natural language narratives for non-expert users. 4) Related IS Theory: Grounded in Decision Support Systems (DSS) theory, this work extends the classical DSS goal. It enhances decision quality not just via predictive accuracy, but by improving the transparency, trustworthiness, and interpretability of the system’s reasoning for stakeholders. 5) Research Method: We conducted an applied, experimental study on a public retail credit dataset. The methodology involved data preprocessing, XGBoost predictive modeling, quantitative evaluation of explanation fidelity with the MEMC metric, and developing a functional web prototype. 6) Summary of Results: The framework effectively identified key credit denial factors with high fidelity, validated by the MEMC metric. The LLM-synthesis layer successfully transformed complex XAI data into clear, understandable, and practical explanations, enhancing the system’s clarity and actionability. 7) Contributions and Impact to IS area: This study contributes a validated framework for building more ethical, transparent, and socially inclusive intelligent systems. Its impact lies in bridging the gap between advanced AI and human-centric requirements, enabling responsible AI adoption and strengthening human-AI collaboration in decision-making.

Ähnliche Arbeiten