Model-Based Trust Analysis of LLM Conversations

LLM-based chatbots are routinely advertised as supporting the collaboration of humans and AI. We study LLM conversations from a knowledge elicitation perspective with the objective of being able to understand and assess the human's trust in knowledge elicited from the LLM and complementary sources. Our approach is supported by the DSML KEML, the Knowledge Elicitation Modeling Language, subject to abstract and visual syntax as well as a model transformation-based model semantics for trust analysis. Conversations are modeled by a combination of sequence diagrams and enhanced argumentation graphs --- the latter for the purpose of relating information pieces (facts and instructions) that are extracted from messages. The analysis of the corresponding models entails trust scores for gathered information (i.e., elicited knowledge).