Project Gaia: Enabling climate risk analysis using generative AI

Project Gaia aims to help analysts search corporate climate-related disclosures and extract data quickly and efficiently using artificial intelligence (AI), particularly large language models (LLMs). Gaia Phase I has surveyed climate risk experts from central banks and supervisory authorities, designed a solution that addresses the requirements articulated by these experts and delivered a proof of concept (PoC) demonstrating the technical feasibility of the concept.

By automating information extraction, the Gaia PoC enabled the comprehensive analysis of climate-related indicators. Furthermore, it offered harmonised metrics despite the heterogeneity of naming conventions and definitions across different jurisdictions. The combination of semantic search together with iterative and systematic LLM prompting enabled Gaia to overcome differences in disclosure frameworks. This offers much-needed transparency and comparability of climate-related information.

Project Gaia integrated LLMs into an application and leveraged it for data extraction. This posed several technical challenges, including LLMs' long response times, randomness (non-repeatability) in their responses and hallucinations. The project report explains a set of concrete design choices that allowed the Gaia PoC to overcome these challenges.

Gaia demonstrates the power of creating AI-enabled intelligent tools to automate existing workflows. For example, macro analysis results presented in this report cover 20 key performance indicators (KPIs) for 187 financial institutions over five years and adding more institutions or KPIs is quick and easy. Thanks to its flexible design, the platform is relevant in a broader context than climate-related data analysis, showcasing the value of AI-enabled applications for central banks and the financial sector. Generative AI promises to change the way we work in the future, and Project Gaia is one of the first comprehensive studies investigating how this can be done in practice.