Geosystems risk and uncertainty: The application of ChatGPT with targeted prompting

Abstract

ChatGPT, a prominent large language model (LLM), is being increasingly used across a wide range of scientific fields. Geosystem engineers and researchers are also posed to leverage ChatGPT to find solutions to challenges encountered in various topics. This study evaluates the accuracy and reproducibility of ChatGPT in responding to different qualitative and quantitative questions, with a particular focus on risk and uncertainty (R&U) in both the Greenfield and Brownfield domains as an important area of interest. The results show the importance of prompting to considerably improve the ChatGPT’s response accuracy and reproducibility. For example, prompting increases the accuracy of responses to qualitative and quantitative questions in the Greenfield domain by 10.4% and 41.8%, respectively. Additionally, prompting enhances the reproducibility of responses, with a 32.1% increase for qualitative questions and a 33.3% rise for quantitative questions in the Brownfield domain. The findings highlight that the greater the comprehensiveness of the prompts, the higher the accuracy and reproducibility of the responses to the questions. The study also acknowledges the potential limitations associated with the sources of information and the contextual influences on the reliability of the response.