A domain specific BERT model for identifying AI disclosures in Islamic banking

This study presents a sentence-level, domain-adapted transformer-based model designed to identify AI-related disclosures within Islamic banking. Although natural language processing (NLP) has advanced rapidly in financial domains, its application to Shariah-compliant financial reporting remains largely unexplored. To address this gap, a decade-long, multi-jurisdictional corpus (2015–2024) of 855 English-language annual reports from 106 full-fledged Islamic banks across 25 countries was constructed. Using a seed-based filtering method, 2,632 sentences were extracted and manually verified, equally balanced between AI and non-AI disclosures. Fine-tuning BERT-base-uncased using the Hugging Face Trainer API yields strong performance (F1-score: 0.9868; Brier Score: 0.0027), indicating robust generalization, reliable calibration, and resilience to noisy, paraphrased, and syntactically varied inputs. Comparative evaluations against logistic regression, Naïve Bayes, Random Forest, and XGBoost demonstrate higher semantic adaptability and clearer interpretability for the transformer-based approach. The core contribution is twofold: (i) the first domain-specific, sentence-level AI-disclosure detector for Islamic banking; and (ii) an openly released corpus, codebase, and model to support regulators, Shariah boards, and policy analysts in monitoring AI adoption and enabling scalable technology auditing within Shariah-compliant finance.