Understanding the Division of Cognitive Effort in Human-AI Interactions

As LLMs are increasingly treated as thought partners that collaborate with users on different cognitive tasks, it is important to understand the division of labor in this interaction: how much cognitive effort is a user offloading to such an AI-based thought partner in each interaction? In this study, we operationalize the amount of cognitive offloading as the time saved when completing a task with AI assistance. We conduct a large-scale study to examine the systematic difference between completing a task independently and completing it with AI assistance across different cognitive tasks. We hypothesize that there is a set of tasks with a slow-down effect – where using AI assistance takes longer than independent completion. Additionally, we compare the time participants predict a task would take to the actual completion time and hypothesize that a calibration error persists in how much offloading AI use affords: users tend to underestimate the time it takes to complete a task with AI assistance while overestimating the time it takes to complete a task independently.