Investigating the Impact of Control in AI-Assisted Decision-Making - An Experimental Study

We ask whether users should adjust to AI systems or vice versa. Levels of automation (LOAs) are task dependent, may vary within one task, and also may change over time. People’s diverse abilities and preferences make the usage of AI systems possibly personal. Automation design is a complicated task. We investigate varying levels of LOAs in one specific decision-making process. For this, we conduct an experiment, where n=24 volunteers participate in a within-subject face-recognition experiment. Face-recognition is an innate ability mastered by humans. Reason are specialized neurological systems. This also makes it an intuitive task. The results show that of the five tested LOAs, each one leads to personal best and personal worst decisions regarding accuracy and time. Similarly, each LOA is preferred or opposed by participants. This shows, that there is no “one-size-fits-all” LOA, suggesting that careful design is required and multiple LOAs should be offered for a task.