Trust-Aware and Explainable Generative AI Frameworks for Interpretable Large Language Models in Critical Applications

Abstract

The rapid adoption of Large Language Models (LLMs) across healthcare, finance, and governance has amplified concerns regarding explainability, accountability, and regulatory compliance. While generative AI systems demonstrate remarkable performance in language understanding and decision support, their opaque architectures and probabilistic reasoning processes hinder trust in high-stakes applications. This paper proposes a comprehensive framework titled ET-GEN (Explainable and Trustworthy Generative Network), designed to integrate interpretability mechanisms, uncertainty quantification, and governance-aligned evaluation metrics within LLM-driven decision systems. The framework incorporates attention attribution mapping, counterfactual reasoning modules, confidence calibration layers, and domain-specific rule alignment constraints. Experimental validation across healthcare diagnosis summarization, financial risk assessment, and public policy classification tasks demonstrates improved interpretability scores, calibrated confidence estimation, and regulatory alignment compared to baseline LLM systems. The results indicate that embedding explainability layers within generative architectures enhances transparency without significantly degrading predictive performance. The proposed model offers a scalable pathway for deploying trustworthy generative AI systems in regulated environments.

Ähnliche Arbeiten