Attributing Mind to Large Language Models: The Effect of Exposure and Individual Differences

Recent developments in large language models (LLMs), have renewed calls to study how people perceive minds in these AI applications. Across 4 experiments that differed in the type of exposure (vignettes or real-time interaction), we found that exposure to LLMs (i.e., Chat GPT, LLaMA, Claude) can increase attributions of mind for both agency (ability to do) and experience (ability to feel). These effects varied as a function of exposure type, with vignettes producing larger effects than real-time interactions, perhaps because users tended to ask fact-based questions during real-time interactions. We also found that individuals who interacted with LLMs more before the experiments, and individuals with a general propensity to anthropomorphize, perceived more mind in LLMs. These findings suggest that as LLMs grow in popularity, and people are exposed to them to a greater extent, the degree to which people attribute qualities of mind to AI systems will also increase depending on the type of exposure. These results pose an intriguing question for future research regarding how long-term exposure or other nuances in the type of exposure to LLMs may influence mind perception.