Context
In the rapidly evolving landscape of artificial intelligence (AI), one of the most promising research directions is the development of systems that can work alongside and collaborate with humans as actual teammates. Effective teamwork is crucial in various fields, such as healthcare, military operations, and disaster response. In these environments, the ability of team members to collaborate can significantly impact outcomes, making the development of socially intelligent AI a vital area of study. Our research contributed to this field by examining how AI can develop an Artificial Theory of Mind (AToM; Williams, Fiore, and Jentsch, 2022) to support and enhance interactions with human counterparts acting as a team.
Collaborative Research Process
Funded by the US Department of Defense program, “Artificial Social Intelligence for Successful Teams” (ASIST; https://artificialsocialintelligence.org/), this large interdisciplinary initiative aimed to develop AI systems that understand and enhance human team dynamics. Our team at the University of Central Florida (UCF) focused on developing new social science theory and concepts that can help AI predict and improve team processes and performance.
The ASIST program involved over 200 scientists from diverse fields, including computer science, psychology, and organizational behavior. A critical aspect of our work was integrating these disciplines to develop AI agents capable of social intelligence. One significant challenge was balancing the technical goals of AI development with the social research goals focused on understanding and improving team behaviors. For instance, allowing natural language communication within our simulated task environment was useful for studying team interactions but presented substantial technical challenges for AI systems. We had to carefully design experiments that allowed for the collection of meaningful behavioral data – communication in general but also parasocial interactions – while ensuring the AI systems could process and respond to this data in real-time (Fiore, Bracken, Demir, Freeman, and Lewis, 2021).
Despite these challenges, the interdisciplinary collaboration within the ASIST program led to innovative solutions (Fiore, Johnson, Robertson, Diego-Rosell, and Fouse 2023). One of the most significant successes was developing a comprehensive testbed using the Minecraft environment to simulate Urban Search and Rescue (USAR) tasks (see video below). This setup allowed us to conduct large-scale experiments, providing valuable data on teams made up of humans with functioning AI. Each research team’s contributions shaped the nature of the experimental team task, the testbed architecture, and the measures of individual and team performance outcomes. This collaboration supported the testing of a host of hypotheses, ranging from expectations of ASI success to the emergence of team leadership and, related to our work at UCF, the potential to profile teams and predict their behaviors during complex collaborations with other humans and with AI.
Results
Central to UCF’s approach in the ASIST project was creating detailed individual and team profiles (Bendell, Williams, Fiore, and Jentsch, 2023), capturing key characteristics within a theoretical framework of teamwork potential and taskwork potential (e.g., spatial abilities, video gaming experience, and social intelligence (see Figure 1). These profiles provided ASI agents with the necessary context to develop an artificial Theory of Mind (AToM), enabling them to make informed and contextually relevant interventions during team tasks (Bendell, Williams, Fiore, and Jentsch, 2023; Williams, Bendell, Fiore, and Jentsch, 2023). The goal of the profiling approach was ultimately to provide the ASI insight into the likely behaviors, struggles, and capabilities of their human teammates so as to tailor their advice and better support the team processes.
Figure 1. The two dimensions of our profiling approach were Taskwork Potential and Teamwork Potential each defined by scores on three measures. Taskwork potential was determined by: spatial ability (SSOD: Santa Barbara Sense of Direction), video gaming experience (VGE: Video Gaming Experience), and a task skills competency test (COMP: Competency test). Teamwork potential was determined by: preferences for team and group work (Collectivism: Psychological Collectivism), attitudes towards social interactions (SD: Sociable Dominance), and social intelligence (RMET: Reading the Mind in the Eyes Test). Team members were categorized as either high or low in potential on each dimension.
Our findings demonstrated that AI agents equipped with social intelligence could significantly enhance team performance, particularly in teams entering the game environment with lower taskwork and teamwork potential. In the USAR missions, AI advisors were able to provide critical support, improving task execution and team coordination. This suggested that AI systems could be particularly valuable in augmenting the capabilities of less experienced or lower-performing teams.
The profiling approach proved effective in predicting team behaviors and outcomes. Our analyses also showed that high taskwork and teamwork potential profiles were associated with better performance as well as more positive perceptions of their AI advisor and the team’s collaborative processes.
Impact and Conclusions
Our research underscores the importance of integrating social intelligence into AI systems. By informing the development of AI agents designed to better understand and predict human behaviors using the profiles, we aimed to create more effective human-machine teams. The study additionally highlighted the importance of interdisciplinary integration whereby collaboration led to the synthesis of concepts from computer science and social science, thus advancing both AI technology and our understanding of human cognition (Fiore et al. 2023).
The relevance of our findings for the future of human-AI teams is significant. AI systems with social intelligence can enhance team performance, particularly in complex environments where effective teamwork is critical. The ASIST program exemplified how bringing together social and computer sciences can enable innovative experiments and yield valuable insights. Further, the program yielded multiple, publicly available datasets and testbeds that may be employed by other researchers to advance our understanding of human-AI teaming (see https://artificialsocialintelligence.org/data/).
Looking ahead, further research is needed to explore the nuances of human-AI interactions. Future studies should examine how different types of AI interventions impact team dynamics and performance and evaluate the long-term effects of AI advisors on team development. This interdisciplinary approach will continue to be crucial as we strive to develop AI systems that are not only technically proficient but also socially aware, paving the way for AI agents that function as true teammates.
References
Williams, J., Fiore, S. M., & Jentsch, F. (2022). Supporting artificial social intelligence with theory of mind. Frontiers in Artificial Intelligence, 5, 750763.
Bendell, R., Williams, J., Fiore, S. M., & Jentsch, F. (2023). Teamwork Traits Associated with Positive Perceptions of the Dependability and Utility of Autonomous Advisors. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67, (pp. 440-445). Los Angeles, CA: SAGE Publications.
Bendell, R., Williams, J., Fiore, S., & Jentsch, F. (2023). Interventions by Artificial Socially Intelligent Agents in Collaborative Environments: Impacts on Team Performance and Knowledge Externalization. In W. Karwowski and T. Ahram (Eds.), Artificial Intelligence, Social Computing and Wearable Technologies: AHFE International Conference, 113, (pp. 165-175). AHFE International Open Access.
Williams, J., Bendell, R., Fiore, S.M., & Jentsch, F. (2023). The Role of Artificial Theory of Mind in Supporting Human-Agent Teaming Interactions. In J. Wright and D. Barber (Eds.), Human Factors and Simulation: AHFE International Conference, 83, (pp. 46-56). AHFE International Open Access. Winner, Best Student Paper Award.
Fiore, S. M., Johnson, M., Robertson, P., Diego-Rosell, P., & Fouse, A. (2023). Transdisciplinary Team Science: Transcending Disciplines to Understand Artificial Social Intelligence in Human-Agent Teaming. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67, (pp. 419-424). Los Angeles, CA: SAGE Publications.
Fiore, S.M., Bracken, B., Demir, M., Freeman, J., Lewis, M. (2021). Transdisciplinary Team Research to Develop Theory of Mind in Human-AI Teams. Proceedings of the Human Factors and Ergonomics Society Annual Meeting (pp. 1605-1609). Los Angeles, CA: SAGE Publications.