Large Language Model-assisted EIA screening: a case study using GPT

Large Language Models (LLMs) have developed rapidly in recent years and are increasingly used for tasks involving the interpretation of human language expressed in text. As many EIA systems rely on EIA screening approaches that are based on the interpretation of descriptive thresholds contained in lists of text, LLMs might hold value for automating aspects of the EIA screening stage. This paper investigates the feasibility of using a customised Generative Pretrained Transformer (GPT) model (a specific type of LLM) as an EIA screening tool. Three versions of a GPT-based screener were developed through an iterative process and tested against 20 real-world EIA cases involving activities regulated by two listing notices under South African law (GNR 983 and GNR 984). The iterative improvement of the model – from GPTv1 through GPTv3—demonstrated improvements in correctly identifying applicable activities that would be triggered. However, the models were not without challenges and specifically struggled with large-scale and highly complicated development proposals involving multiple triggers. The results demonstrate the potential value of GPTs but also highlight the importance of human oversight and the need for iterative refinement tailored to specific contexts.