Toward a post-humanist pragmatics: investigating anaphoricity in human-AI interactions

Abstract Focusing on anaphora, the paper draws attention to the need to test pragmatic theories in light of the rise of post-humanist communicators, i.e. chatbots. The paper analyzes data collected from human-AI interactions including different types of anaphora. The analysis shows that chatbots perform well in recovering the antecedents of sentential and cross-sentential anaphors. The only exception is Copilot which seems to struggle with cross-sentential anaphora. However, all of the chatbots investigated could not recover the antecedents of all cross-speaker anaphors. This makes AI-generated responses less cohesive and, consequently, less coherent. The results suggest that these chatbots find discourse anaphora much more difficult to process than their sentential counterparts due to the context-sensitivity of the former. The paper indicates that AI-generated texts exhibit lower levels of anaphoricity than human-authored texts. Although this should not be interpreted as a deterministic rule for detecting AI-generated content, it indicates that the different types of anaphora may constitute a probabilistic feature among many in multi-dimensional detection frameworks.