Enterprise AI Analysis

Teaching GPT to Explain Itself: An Interpretable AI Framework for Rejecting Loan Applications in Regulated Finance

In regulated financial scenarios, the "black-box" nature of GPT-style large models conflicts sharply with the stringent compliance requirements for loan rejection explanations—existing methods either fail to interpret unstructured risk signals (e.g., textual claims of "impending job changes" in applications) or lack regulatory alignment and traceability, hindering large-scale AI deployment. To address this gap, this study proposes an interpretable AI framework enabling GPT to self-explain loan rejections, centered on three core innovations: 1) regulatory-aligned lightweight fine-tuning via LoRA, which injects multi-regional regulatory rules (EU GDPR, China's Guidelines for AI in Banking) and bank internal policies into GPT without full-model retraining; 2) standardized interpretation of unstructured factors, converting textual risk signals into structured outputs with "factor description-risk weight-regulatory basis-data source"; 3) end-to-end audit closure, tracking the entire "factor extraction-decision generation-explanation output" process to meet regulatory penetration testing requirements. Validated on 200,000 desensitized personal consumer loan datasets from Bank of Hangzhou (2021-2024), the framework achieves an AUC-ROC of 0.93 (matching the prediction performance of basic GPT) while reaching a 92% explanation compliance rate and 98% traceability completeness. It reduces explanation generation time from 15 minutes (manual process) to 0.3 seconds, and lowers loan rejection complaint rates by 40%. Theoretically, this study fills the gap between large language model interpretability (XAI) and financial compliance; practically, it provides a low-cost, scalable solution for financial institutions of all sizes, with its logic extendable to insurance underwriting and securities risk control.