Large Language Models (LLMs) have evolved into conversational partners with individuals across contexts. Importantly, LLMs are potent persuasive tools, shaping human attitudes and behaviors. We ask whether LLMs can shape basic personal values and value-based behavior.
We formed AI agents with a value-based persona, focusing on either openness to change or conservation values. N = 358 participants were randomly assigned to one of four conditions: bot persona (creative vs. conservative), and topic of conversation (promoting freedom of speech vs. safety of users on social media). Participants answered questionnaires about values and attitudes, chatted with the bot for 5 minutes, followed by additional questionnaires, a short essay about the assigned topic, and a creative alternative uses task.
Self-reported, abstract values did not change following the interaction, exemplifying the stability of personal values. Nevertheless, AI-based analysis of the interaction indicated that individuals' chat expressed values that match the bot persona and topic. More importantly, the effect carried over past the interaction, with following essays expressing values and written in a style congruent with the bot persona. Last, performance in the creativity task differed by both bot persona and initial values of the participants. Individuals who encountered a bot whose persona matched their values showed appropriate creativity (low for conservatives, high for open). Those who encountered a bot whose persona conflicted with theirs were consistently more creative.
Past research indicates that AI systems are not neutral in their values but carry inherent biases. Our results suggest that even brief interactions with a biased AI system may carry effects into human behavior over time and into unrelated contexts. Although individuals did not experience a conscious effect of value change, we saw changes in their expressed attitudes, writing style, and even creative behavior. These results have far-reaching consequences for the integration of AI in human societies