Algorithmic Accountability in Sustainable Finance: A Systematic Review of AI, NLP, and Machine Learning Integration in Global ESG Risk Evaluation

Over the past decade, ESG criteria have moved from the periphery of investment analysis to its centre, yet the infrastructure supporting ESG evaluation has not kept pace. Rating agencies disagree with each other, companies report what suits them, and the smaller businesses that make up the vast majority of the global economy are effectively shut out of sustainable finance markets entirely. This paper examines how AI, Natural Language Processing (NLP), and machine learning are being deployed to address these gaps, drawing on a systematic review of eleven empirical studies. The findings are, in places, counterintuitive. Simple Linear Regression outperforms LSTM networks and BERT on corporate ESG text corpora — achieving 68.09% accuracy compared to 25.53% and 46.81% respectively — because corporate sustainability reports are so formulaic that deep learning models overfit to noise rather than signal. Where AI does deliver clear operational gains is inside structured enterprise architectures, where a 40% reduction in manual processing time has been documented. This review synthesizes these findings into a proposed algorithmic accountability architecture aligned with ISSB convergence efforts, and maps a realistic path toward democratising ESG compliance across firm sizes and geographies.