SAFE Artificial Intelligence to Improve Environmental, Social, and Governance Sustainability

ABSTRACT Environmental, social, and governance (ESG) sustainability is a key dimension for achieving several sustainable development goals (SDGs). While most existing studies focus on large listed firms and treat ESG scores as continuous outcomes, there is still limited empirical evidence on small and medium enterprises (SMEs), despite their central role in the European economy. This paper addresses this gap by proposing a framework for predicting ESG ratings that explicitly accounts for their ordinal nature. In addition, we introduce a novel evaluation framework grounded in safe and ethical AI principles, designed to assess the trustworthiness of model predictions. The empirical analysis is conducted on a dataset of Italian SMEs, a segment largely underrepresented in the ESG literature. Our results show that Ordinal Forest models better exploit the ordinal structure and provide more stable predictions, while Ordinal Logistic Regression achieves higher accuracy but lower robustness. Overall, the findings highlight both the importance of modelling ordinality and the challenges associated with predicting ESG outcomes in the SME context.