Invited Talk by Yu Wang on Tuesday, May 05, 15:00

To enhance the trustworthiness of Large Language Models (LLMs), there are two areas we should continue to work on. The first is the development of robust methods that make users aware of LLM involvement in daily life, such as LLM-based explanations. Human and machine-text detection method, with sufficient ethic consideration, can allow users to determine the degree to which content should be trusted. The second is involving human users in the co-construction of explanations for LLM outputs. In my presentation, I will emphasize the first area and present my previous and current work on revealing the differences between human language and model-generated language. First, I hypothesize that human language is constrained by cognitive resources, which creates distinct patterns that differentiate it from model-generated text. Based on this idea, a method called FourierGPT is developed to detect model-generated text. In follow-up work, I examine how human–human dialogue and human–LLM dialogue differ at the syntactic level, and I use these syntactic features to detect LLM-involved dialogue. Finally, I present an analytical study showing that the periodicity of information in human and LLM-generated text differs significantly.