Dr Arka Ujjal Dey

Firms in global supply chains can strategically exploit information disclosure to amplify positive ESG commitments while concealing poor environmental and social performance through low-visibility channels. Such strategic distribution of information creates information asymmetries in which stakeholders cannot always distinguish genuinely responsible organisations from those making unsubstantiated claims. No automated, scalable mechanism exists to verify the consistency of corporate disclosure. We address this gap by proposing a retrieval-augmented large language model (LLM) verification framework. This approach adapts computational fact-checking techniques, originally used for social media misinformation detection, to the domain of corporate ESG disclosure. The framework classifies claims against independently cross-validated disclosure data, so that what a firm asserts publicly can be automatically compared with what its disclosures substantiate. We validate our approach using ESG disclosure microdata covering global clothing brands and their supply chains. A disclosure-recovery benchmark of 34,572 brand–criterion judgments shows that brand websites surface audited promotional disclosure at two to three times the rate of operational disclosure. A verification study of 606 promotional claims finds that only 5–7% can be substantiated against audited evidence, with results stable across four open-weight language models. The framework offers regulators, consumers, and supply chain partners alternative audit mechanisms to detect claim-disclosure gaps, reducing information asymmetries that facilitate deceptive communication strategies.

Fact-checking systems have gained traction as scal-able solutions, yet they often face challenges such as handling diverse evidence sources, integrating multimodal data, and presenting comprehensive narratives. In this work, we propose CRAVE (Cluster-based Retrieval Augmented Verification with Explanation), a novel framework that integrates retrieval-augmented Large Language Models (LLMs) with clustering techniques to address multimodal misinformation on social media. The framework is designed to process multi-modal inputs (text and images) and iteratively refine evidence through agent-based mechanisms. We validated the framework on multiple real-world and synthetic datasets, showing that breaking up evidence into narrative clusters improves both retrieval precision, clustering quality, and judgment accuracy, showcasing its potential as a robust decision-support tool for fact-checkers.