Application of Signal Detection Theory in Evaluating Trust of Information Produced by Large Language Models

Abstract

Hallucinations are a major threat in applying generative artificial intelligence technologies like large language models (LLMs) to high-consequence domains like nuclear electricity generation. Inappropriately trusting an LLM can have deleterious consequences, such as misusing unreliable LLMs and disusing reliable LLMs. Identifying methods to evaluate the trustworthiness of LLMs and resultant user trust is an open area that demands future research from both a technical perspective and a human factors perspective. In this paper, we highlight the challenges in evaluating trust in LLMs and then introduces SDT as a potential framework that may overcome these challenges. We then propose using signal detection theory (SDT) as an evaluative framework for comparing trust and trustworthiness when interacting with LLMs in high-consequence domains like nuclear electricity generation.