Supporting Social Interactions In Human-Ai Teams: Profiling Human Teammates From Sparse Data

Abstract

Artificial intelligence has been developed to perform all manner of tasks but has not gained capabilities to support social cognition. We suggest that teams comprised of both humans and artificially intelligent agents cannot achieve optimal team performance unless all teammates have the capacity to employ social-cognitive mechanisms. These form the foundation for generating inferences about their counterparts and enable execution of informed, appropriate behaviors. Social intelligence and its utilization are known to be vital components of human-human teaming processes due to their importance in guiding the recognition, interpretation, and use of the signals that humans naturally use to shape their exchanges. Although modern sensors and algorithms could allow AI to observe most social cues, signals, and other indicators, the approximation of human-to-human social interaction -based upon aggregation and modeling of such cues is currently beyond the capacity of potential AI teammates. Partially, this is because humans are notoriously variable. We describe an approach for measuring social-cognitive features to produce the raw information needed to create human agent profiles that can be operated upon by artificial intelligences.

Factor Analysis of a Generalized Video Game Experience Measure

“…Findings show that the VGEM taps into general experience with respect to duration, frequency/intensity, and self-reported skill. These features of video game experience may transfer to novel task performance (Bendell et al, 2021a). With the VGEM, researchers can use its assessment to help account for the variance of prior experience in video games as it can influence the performance of participants in research and training.…”

“…Accounting for variance allows researchers to better understand differences in task performance that are more likely due to video game experience. Further, by understanding participant characteristics, these kinds of measures can be used to create profiles of players (Bendell et al, 2021a), which can be used both in gamified research tasks to understand performance, as well as to provide a priori information to artificial intelligence that would aid the development of an Artificial Theory of Mind, allowing the AI to better interpret and predict human behaviors (Bendell et al, 2021b; Williams et al, 2022). Overall, the VGEM can help guide researcher decisions whenever gamified experimental tasks are used.…”

“…The video game experience measure contained a total of 24 items, though only 3 of these were removed as they were strictly capturing Minecraft-specific video game experience instead of more generally applicable video game experience that could be employed to assess video game experience in other contexts. It is noted, however, that specific video game experience can be a useful in determining the level of a participant’s experience (see Bendell et al, 2021a). Participants completed the video game experience as part of a set of surveys during the first phase of the study.…”

“…Because of the limitations of current methods for measuring a participant’s video game play experience (for further discussion on this, see: Bendell et al, 2021a; Williams et al, 2021) in a way that can account for performance differences, and that has been empirically validated (Abeele et al, 2020), we developed a tool that could potentially fill this gap. This was done in the context of a larger research program, the Defense Advanced Projects Research Agency’s “Artificial Social Intelligence for Successful Teams” (ASIST; DARPA, 2019).…”

Transdisciplinary Team Science: Transcending Disciplines to Understand Artificial Social Intelligence in Human-Agent Teaming

“…These systems vary in complexity, ranging from dyadic interactions between a human and AI, collaborative interactions, where AI is a member of the team, to more facilitating or coaching interactions where AI is guiding teams. Across these, researchers must examine, not simply how AI alters completion of tasks, but also how AI affects human team members and their interactions (Bendell, Williams, Fiore, & Jentsch, 2021a; Kohn, de Visser, Wiese, Lee, & Shaw, 2021; Schelble, Flathmann, McNeese, Freeman, & Mallick, 2022; Terblanche, Molyn, de Haan, & Nilsson, 2022).…”