Linguistic perspective-taking in verbal communication with LLMs: evidence of an egocentric bias

Abstract

Since the expansion of generative AI’s functions, chatbots such as Open AI’s Chat-GPT appear as more than tools. They can be perceived as communicative partners, shaping exchanges with high variability and linguistic competence. Current versions, such as Chat-GPT 5.5, even provide spoken language using different verbal styles and vocabulary, intonation and prosody. Investigating this new context of communication, we question whether people expect chatbots to detect non-literal meanings. Extending the illusory transparency of intention research paradigm introduced by Keysar, information was presented through scenarios of interactions with the Chat-GPT voice function that disambiguated the presented verbal information as either non-literal (sarcastic) or literal (sincere) in an within-subject experimental design (N = 327). Scenarios were placed within the higher education context. Results reveal that participants did not only falsely expect the chatbot to detect a non-literal meaning, but they also even expected the chatbot to reply accordingly. Implications for higher education contexts are discussed.