Testing the Machine-Likeness Hypothesis: How Humans Communicate Like Intelligent Agents

The emergence of human-like intelligent conversational agents has led researchers to consider how intelligent agents (IAs) could influence human communication and further challenge human civilization. To address these questions, the study proposes the machine-likeness hypothesis, which argues that communicating with IAs could make human communication more machine-like. The study analyzed a large-scale dataset of human conversations with ChatGPT for a year. By comparing it with a corpus of simulated human-ChatGPT conversations based on real human-human conversations, the study found that the difference in perplexity between human- and ChatGPT-generated messages is smaller in real conversations, and the difference decreased faster in the non-task-oriented conversations over turns, suggesting a machine-like communication style. The findings extend conversational alignment studies by examining how humans become machine-like through alignment and challenge the technology determinism on the impacts of machines on human communication, as the alignment to IAs may work in proactive ways.