Understanding Social Influence on Decision-Making in Human-AI Teams

Abstract

This research investigated social influence on decision-making in human-AI teams, with a focus on two aspects we identified as particularly important: social conformity and authority influence. Our social conformity research consists of two studies. The first study examined factors driving social conformity: informational influence (the desire to obtain information from others) and normative influence (the desire to seek social approval and avoid punishment from others) in human-AI teams. The results indicated that humans and AI can exert a comparable level of informational influence, but the normative influence in human teams is stronger than in AI teams. In the second study on social conformity, we varied the probabilistic decision accuracy of the information presented. This allowed us to observe how individuals use majority rules and probabilistic decision accuracy in their decision-making processes within human-AI teams. We discovered that people tend to combine both strategies when making decisions. However, this combination may result in information overload, leading to sub-optimal decision outcomes. Our authority influence research encompassed further two studies to investigate the drivers of authority influence: formal authority (driven by social roles and their associated powers) and real authority (driven by the rights to determine final outcomes) on decision-making in human-AI teams. In the first study, we manipulated how the formal authority of human and AI advisors was presented. The results indicated that formal authority alone does not significantly impact people’s decision-making whether an advisor is human or AI. In the second study, we manipulated real authority by granting an authority figure the power to determine whether to punish participants. We found that real authority can significantly impact people’s decision-making in human-AI teams, causing people to make sub-optimal decisions. Moreover, we found that humans and AIs can exert a similar degree of real-authority influence. This research both enriches the theoretical understanding of social influences in human-AI teams and provides insights for designing practical guidelines for enhancing human-AI collaboration.

Date of Award	2 Oct 2024
Original language	English
Awarding Institution	University of Bath
Sponsors	UKRI EPSRC
Supervisor	Janina Hoffmann (Supervisor) & Eamonn O'Neill (Supervisor)