Trust plays a critical role in both effective teamwork and successful integration of AI into human-AI teams (HATs), which are becoming increasingly interconnected to tackle large-scale problems. As these HATs grow in complexity, so do the challenges of building, maintaining, calibrating, and aligning trust across HATs, as trust is dynamic, socially influenced, and capable of spreading and evolving over time. To understand the underexplored question of how trust and distrust can spread across HATs, we conducted a laboratory experiment with two interconnected HATs (N=40) distributed across the U.S., each comprising two humans and one AI collaborating over a remotely piloted aircraft system (RPAS) conducting simulated reconnaissance tasks. Our findings show that distrust toward the AI teammate spread more easily than trust, both within and across HATs. While intra-HAT (dis)trust in the AI developed through direct interaction, inter-HAT (dis)trust was more vulnerable to word-of-mouth distrust. Additionally, participants’ trust in the (dis)trust spreader, as well as their team-level trust, reflected in-group favoritism and out-group bias. This work presents the first empirical investigation into trust and distrust contagion across HATs, expanding the current understanding of trust dynamics in multi-agent and multi-team systems. It identifies potential mechanisms driving trust and distrust transfer between interconnected HATs and explores the individual, team, and multi-team level impacts. In doing so, we lay the groundwork for developing theories on trust contagion in multi-agent, multi-team systems, providing valuable insights for future research and design strategies to optimize human-AI collaboration in complex environments.

The success of human-AI teams (HATs) requires humans to work with AI teammates in trustful ways over a certain time period. However, how trust evolves and changes dynamically in response to human-AI team interactions is generally understudied. This work explores the evolvement of trust in HATs over time by analyzing 45 participants' experiences of trust or distrust in an AI teammate prior to, during, and after collaborating with AI in a three-member HAT. Our findings highlight that humans' expectations of AI's ability, integrity, benevolence, and adaptability influence their initial trust in AI before collaboration. However, this initial trust can be maintained or revised through the development of situational trust during collaboration in response to the AI teammate's communication behaviors. Further, the trust developed through collaboration can impact individuals' subsequent expectations of AI's ability and their collaborations with AI. Our findings also reveal the similarities and differences in the temporal dimensions of trust for AI and human teammates. We contribute to CSCW community by offering one of the first empirical investigations into the dynamic and temporal dimension of trust evolvement in HATs. Our work yields insights into the pathways to expanding the methodological toolkit for investigating the development of trust in HATs, formulating theories of trust for the HAT context. These insights further inform the effective design of AI teammates and provide guidance on the timing, content, and methods for calibrating trust in future human-AI collaboration contexts.

This study investigated how people coordinate their decisions and actions with a risky or safe computer-generated agent in a humanoid or non-humanoid form and how this experience influences later behavior when acting alone. In Experiment 1, participants first repeatedly crossed continuous traffic in a virtual environment with a humanoid computer-generated agent (Figure 1). Participants were specifically instructed to cross with an agent that was programmed to be either safe (taking only large gaps) or risky (also taking relatively small gaps). Participants then repeatedly crossed the same roadway alone. We found that participants’ experiences with crossing safe vs. risky gaps with an agent persisted in later trials when the participants crossed alone, such that participants accepted tighter gaps if they were previously paired with a risky than a safe agent. In Experiment 2 (Figure 2), we tested whether experience crossing with a risky or safe non-humanoid object (a floating box) also influenced later behavior when crossing alone. We again found that participants who crossed with the risky object partner took tighter gaps when later crossing alone than those who crossed with the safe object partner. The Discussion focuses on the impact of experiences with virtual agents on perception–action tuning and the potential of using virtual agents for training safe road-crossing behavior.