Analyzing climate risk disclosures in the banking sector using transformer models and generative AI

Climate-related financial disclosures remain fragmented and often lack forward-looking clarity, limiting their usefulness for risk assessment and decision-making. This study proposes a deep learning–based framework that integrates ClimateBERT, a domain-specific transformer model, with GPT-4o (Generative Pre-trained Transformer 4 Omni), a multimodal generative artificial intelligence model developed by OpenAI, to systematically analyze climate disclosures from 162 English-language reports issued by 17 commercial banks in Thailand between 2019 and 2023. The approach classifies textual data across five dimensions: climate relevance, sentiment, commitment, specificity, and alignment with the Task Force on Climate-related Financial Disclosures (TCFD) framework. The results show that climate-related disclosures account for approximately 14–21% of total content and have generally increased over time. Around 50% of climate-related text reflects commitment, of which 80–90% relates to current or completed actions, indicating limited forward-looking disclosures. Sentiment analysis reveals that opportunity-related narratives slightly exceed risk-related ones, with a strong emphasis on product and service innovation. In terms of specificity, only about 25% of disclosures provide detailed information. The classification models demonstrate high reliability, achieving accuracy rates between 95% and 100% across all dimensions. By combining transformer-based models with generative AI, this study advances methodological approaches to analyzing unstructured sustainability disclosures. It also provides practical insights for improving transparency, encouraging forward-looking reporting, and supporting regulatory development aligned with TCFD recommendations. The proposed framework is scalable and can be applied to other sectors and geographical contexts.