Publication

Abstract : Are individuals capable of recognizing misleading advice? Does the AI assistance delivery mechanism impact user reliance and decision-making? We conduct a large-scale, randomized human subject study to gauge how different AI assistance delivery mechanisms affect decision-making processes. We find that subjects who received assistance only on request demonstrated a more critical engagement with the advice and were less likely to be misled compared to subjects who received unsolicited assistance. This finding has large downstream consequences as unsolicited auto-complete is the normal user experience for many AI applications. We provide actionable insights for designing AI tools, emphasizing the need for mechanisms that support user autonomy in order to increase cognitive engagement. We make our experimental platform open-source to support future research on AI-human interaction and decision-making.