Researchers have developed a specialized natural language processing model designed to detect artificial intelligence disclosures within the annual reports of Islamic banks. The study, published in Discover Analytics, introduces FinAI-BERT-IslamicBanks, a domain-adapted transformer-based classifier that operates at the sentence level. This approach addresses a significant gap in financial natural language processing, where previous models have largely focused on conventional finance or general environmental, social, and governance contexts, leaving the unique linguistic and regulatory paradigms of Shariah-compliant banking largely unexplored.
The core of the study involves the construction of a comprehensive, multi-jurisdictional corpus to train and evaluate the model. The team compiled 855 English-language annual reports from 106 full-fledged Islamic banks across 25 countries, covering a decade from 2015 to 2024. To create a balanced dataset for training, the researchers employed a seed-based filtering method to extract sentences containing AI-related terminology. Following manual verification to ensure contextual relevance, the final dataset comprised 2,632 sentences, equally divided between 1,316 AI-related disclosures and 1,316 non-AI sentences. This balanced structure was critical for minimizing class imbalance during the model training process.
The model architecture is based on BERT-base-uncased, a pre-trained transformer model known for its ability to capture bidirectional context in text. The training process involved stratifying the dataset into an 80% training set and a 20% validation set. Sentences were tokenized, lowercased, and padded or truncated to a maximum sequence length of 128 tokens to accommodate the model’s input requirements. The training was conducted on a Google Compute Engine GPU backend using mixed-precision arithmetic to enhance computational efficiency.
During the fine-tuning phase, the model was trained for three epochs with a linear learning-rate schedule initialized at 2 × 10⁻⁵ and a weight decay of 0.01.
The performance of FinAI-BERT-IslamicBanks was evaluated using standard classification metrics. The model achieved an accuracy of 98.67% and an F1-score of 0.9868, indicating a high capability to differentiate between AI and non-AI sentences. The validation loss was recorded at 0.0650, which, coupled with the high accuracy, suggests strong generalization to unseen data. Furthermore, the model attained a Brier Score of 0.0027, a measure of probabilistic calibration. This exceptionally low score confirms that the model’s predicted probabilities are highly reliable, a crucial attribute for applications in regulatory auditing where confidence in predictions is paramount.
In a detailed confusion matrix analysis, the model correctly identified 262 out of 263 non-AI sentences and 263 out of 264 AI-related sentences in the validation set. This resulted in only one false positive and one false negative, demonstrating near-perfect classification. The Receiver Operating Characteristic curve analysis revealed an area under the curve of 0.9999, affirming the model’s strong discriminative capacity across a wide range of threshold values. These results indicate that the model can effectively maintain both sensitivity and specificity, which is essential for parsing nuanced AI discourse within highly domain-specific financial narratives.
To assess the model’s robustness, the researchers conducted a stress test using a synthetic suite of 40 manually constructed examples. These examples were categorized into edge cases, noise-injected sentences, syntactically altered text, and translated variants. The model demonstrated resilience, achieving 90% accuracy in both the translated and noise-injected categories. It also maintained 80% accuracy in edge cases and syntactically altered examples. This performance suggests that the model is stable when confronted with diverse, non-canonical sentence constructions, a critical feature for real-world application where financial reporting styles vary significantly across jurisdictions.
The study also benchmarked the transformer-based model against four traditional machine learning classifiers: Logistic Regression, Naive Bayes, Random Forest, and XGBoost. These baseline models were evaluated using TF-IDF representations of the same dataset. While Random Forest and XGBoost achieved a high accuracy of 99.05%, they relied heavily on lexical patterns and exhibited limited robustness to linguistic variation. In contrast, the transformer-based model, with an accuracy of 98.86%, demonstrated superior semantic generalization and cross-case stability. Statistical tests, including McNemar’s test, showed that the transformer model had a superior error profile, with fewer false positives and no false negatives compared to the TF-IDF baseline.
The researchers also investigated potential bias in the model’s predictions relative to sentence length. By plotting predicted AI probabilities against the word count of sentences, they found a weak positive correlation with a coefficient of 0.20. This indicates that longer sentences are only slightly more likely to be classified as AI-related. The analysis confirmed that model confidence remains relatively stable across varying sentence lengths, implying that the model is not significantly skewed by verbosity. This robustness to sentence length is a desirable trait for analyzing diverse narrative structures found in annual reports.
Despite the strong performance, the authors acknowledge several limitations of the study. The annotated dataset of 2,632 sentences is considered modest for transformer training, which may risk sparse coverage of rare edge cases. The reliance on English-language reports restricts the model’s generalization to other languages, such as Arabic or Malay, which are prevalent in many Islamic banking jurisdictions. Additionally, the high performance may partly reflect a distributional fit or latent reliance on specific AI keyword features. The authors suggest that future work should include keyword-masked evaluations and out-of-vocabulary saliency reviews to further bound the risk of overfitting.
The study contributes to the field by providing the first domain-specific, sentence-level AI-disclosure detector for Islamic banking. It also releases an openly documented pipeline, corpus annotations, and trained model weights to support reproducible research. These resources are intended to assist regulators, Shariah boards, and policy analysts in monitoring AI adoption and enabling scalable technology auditing within Shariah-compliant finance. By mapping Shariah principles to AI governance pillars such as transparency and explainability, the model offers a framework for assessing the quality of AI disclosures in this specialized financial sector.
The practical applications of this model extend beyond academic research. It can serve as a screening tool for regulatory bodies to monitor AI-related initiatives disclosed in formal reporting. The extracted AI sentences can be used to construct disclosure indices or country-level readiness benchmarks, providing novel inputs for policy analysis and cross-jurisdictional comparison. The modular nature of the model allows for lexicon expansion and dynamic fine-tuning with emerging terminology, paving the way for future research in multilingual Islamic finance corpora. This work underscores the growing importance of domain-adapted NLP tools in ensuring transparency and ethical compliance in the digital transformation of financial services.
Beyond empirical benchmarking, the research articulates an underlying governance architecture connecting Islamic jurisprudence with corporate technology reporting. The author aligns core Shariah principles, including the realization of public interest (maslaḥah), the fulfillment of Islamic objectives (maqasid al-Shariah), and the avoidance of harm and contractual uncertainty (ḍarar and gharar), with recognized artificial intelligence governance pillars. Under this conceptual mapping, rigorous reporting requires financial institutions to disclose their operational boundaries, data provenance, systemic risks, human intervention mechanisms, and formal Shariah supervisory oversight pathways, ensuring that algorithmic adoption respects established ethical doctrines.
The document preparation and extraction protocol also incorporated precise linguistic filtering steps. When processing the multi-jurisdictional annual reports, the pipeline removed non-substantive textual noise, including document headers, footers, tables of contents, formal legal disclaimers, and repetitive boilerplate text. Candidate sentences were flagged through an expert-curated seed vocabulary capturing foundational and emerging technologies, such as robotic process automation, predictive analytics, algorithmic decision-making, natural language processing, and machine learning. Non-AI sentences were then randomly sampled from the unflagged text before comprehensive manual review confirmed the contextual validity of each item in the balanced corpus.
Complementary statistical diagnostics provide further insight into the classifier’s performance envelope. On the paired test set, the domain-adapted model attained an accuracy of 0.987 with a 95 percent confidence interval spanning 0.973 to 0.994, alongside an identical F1-score of 0.987 with a 95 percent confidence interval of 0.976 to 0.996. While comparative evaluations confirmed a superior error distribution over lexical baselines, DeLong’s test indicated no statistically significant difference in discriminative area under the curve with a delta AUC of 0.011 (z = 0.307, p = 0.759), reflecting near-ceiling ranking performance across tested model configurations.
To address remaining technical risks, the study outlines targeted validation and expansion pathways. To confirm that predictions reflect genuine semantic context rather than superficial vocabulary overlap, future implementations can incorporate explainability techniques like SHAP and LIME to evaluate token-level attributions. Furthermore, researchers propose expanding institutional coverage through bank-level and time-split holdout sets, domain-adaptive pretraining across complete unsegmented reports, and multilingual extensions into Bahasa and Arabic financial corpora.
Subject of Research: Development and evaluation of a domain-specific BERT model for detecting AI disclosures in Islamic banking.
Article Title: A domain specific BERT model for identifying AI disclosures in Islamic banking
Article References: Zafar, M. B. (2026). A domain specific BERT model for identifying AI disclosures in Islamic banking. Discover Analytics, 4(1), Article 12. https://doi.org/10.1007/s44257-026-00061-w
Image Credits: AI Generated
DOI: 10.1007/s44257-026-00061-w
Keywords: Islamic banking, BERT, Natural language processing, AI disclosure, Financial reporting, Transformer models, Shariah compliance, Machine learning, Text classification, Regulatory auditing, Financial NLP, Sentence-level analysis, Model calibration, Open source AI, Banking technology
News Source: Denise Maddox. (November 13, 2026). BERT model identifies AI disclosures in Islamic banking reports. Scienmag.
