Have a Heart: Predictability of Trust in an Autonomous Agent Teammate through Team-Level Measures of Heart Rate Synchrony and Arousal

Abstract

Progression toward sophisticated machines with the capacity to act as partners in tactical and strategic situations means that human operators will increasingly rely on collaborative input from agent teammates (e.g., Masiello, 2013). However, plans to team autonomous agents with humans raise new questions regarding the effects that such teammates might have on important team psychological processes, such as team cognition and trust. Specifically, it is not known how modifications in team structure, such as changes in team size, influence team dynamics and psychological processes when the team includes an artificial agent, nor how trust established in such teams transfers to new environments, nor how measures that have been used to predict trust in humans generalize to agent teammates. Our current research examined these effects through the detection and analysis of an objective team phenomenon known as physio-behavioral coupling (PBC) using Multidimensional Recurrence Quantification Analysis (MdRQA; Wallot, Roepstorff, and Mønster, 2016) of shared physiological arousal during initial team formation and training. In particular, as shared physiological arousal measured in changing heart rhythms within human teams has been shown to be associated with measures of trust (Mitkidis, McGraw, Roepstorff, & Wallot, 2015) and concern for others (Konvalinka et al., 2011), we investigated how shared physiological arousal predicts willingness to trust an agent teammate in a novel task environment. We conducted an experiment consisting of collecting physio-behavioral data (i.e., heart rate) from teams of different sizes as they performed a series of collaborative, consensus building tasks. The independent variable was team size (teams of 2 or 3 human players, with an artificial agent teammate always present), and there were two separate team-oriented tasks: A first-round consensus-building wagering task, and a second-round task in which teams were able to make wagers on the expected performance of the agent teammate in a subsequent maze running task called Checkmate (Alarcon et al., 2017). We predicted that complimentary combinations of PBC (e.g., measures of overall similarity and stability in heart rate dynamics) obtained from MdRQA, along with self-reported measures of team and agent trust, would be positively related to future trusting behaviors in the agent teammate, and that increasing the number of teammates would result in higher order, more complex structure in the physio- behavioral data that would not be reducible to simpler patterns (e.g., Wallot et al., 2016). To this end, we predicted that measures of self-reported trust and multivariate PBC would be reducible to meaningful lower dimensional structures using principal components analysis (PCA), and that PBC calculated from the first task from the full team, but not from averages aggregated from subsets of the team, would significantly predict trusting behavior in the second task. Ninety-two participants (31 men and 61 women) recruited from the campus of a midwestern university in the U.S. took part in this study (19 dyads and 18 triads). Ages ranged from 18 to 42 ( M = 22, SD = 5.48). The experiment was a univariate (team size; two or three human teammates with an agent teammate always present) between-subjects design. Self-reported measures were collected from each team member before each of the two tasks and included items that measured: Team ability, team benevolence, team integrity, and team trust (adapted from Mayer & Davis, 1999); trust in human teammates (adapted from Naquin & Paulson, 2003); agent competence, cognitive trust in the agent, emotional trust in the agent, intention to delegate to the agent, and intention to adopt the agent as an aid (adapted from Komiak & Benbasat, 2006); and collective efficacy (adapted from Riggs & Knight, 1994). Factor analysis of the composite scales from aggregated survey data indicated the data loaded well onto factors that corresponded to trust in the team and trust in the agent teammate. Factor analysis of MdRQA from the full team and from the averaged lower order analyses showed that each had one component with an eigenvalue greater than what would be expected by chance. Results from analyses using logistic regression to predict Checkmate betting showed that self-reported measures of trust in the agent and MdRQA of full team PBC in the initial task significantly predicted subsequent trusting behavior in an agent teammate in Checkmate, but lower-order PBC estimated from averages of team subgroups did not. These results suggest that multivariate team-level coupling has predictive power in subsequent team outcomes that cannot be fully captured using data aggregated from subgroup averages, and that measures of PBC measured from human teammates is related to trust in an agent teammate. We note two important contributions of the present study. First, that PBC and subjective measures of trust were significant predictors of observed trusting behavior regardless of team size suggests that important team processes and outcomes are at least partially invariant to changes in team size, a promising outcome for the prospect of meaningfully scaling measures of PBC beyond the typical dyadic context. Second, we have shown that shared team- level arousal is a significant predictor of subsequent trusting behavior in an agent teammate in a novel task, demonstrating that these objective measures are extensible to trust in non-human partners.

In the Rough: Evaluation of Convergence Across Trust Assessment Techniques Using an Autonomous Golf Cart

“…As such, we chose to use subjective selfreport, behavioral, and physiological data to assess trust, each of which has a body of literature detailing its use for assessment of trust (Ajenaghughrure et al, 2020;Akash et al, 2018;Banks et al, 2018;Mitkidis et al, 2015;Sauer et al, 2016;Schwarz et al, 2019;Thayer et al, 2012;Tolston et al, 2018;Wang et al, 2018;Figure 8. Scatterplots for physiological measures.…”

“…The AICP-R consists of 10 questions evaluated using a 5-point Likert scale ranging from "strongly disagree" to "strongly agree." Past research has suggested that HRV and HR may increase when a user's workload is higher due to lower trust resulting in a greater degree of monitoring of the system (Khalid et al, 2016;Mitkidis et al, 2015;Perello-March et al, 2022;Tolston et al, 2018). Additionally, these measures are indications of stress which may be caused by decreased levels of trust (Reimer et al, 2010(Reimer et al, , 2016.…”

“…While a system could routinely ask its user to evaluate his/ her trust, there are potential methodological concerns (e.g., anchoring and response bias) in addition to concerns over whether such a redundant request may be seen as an annoying behavior which could impact the use of such a system (Segura et al, 2012). In an effort to develop a real-time assessment of trust, research has examined the utility of physiological and behavioral measures such as heart rate (Khalid et al, 2016;Mitkidis et al, 2015;Perello-March et al, 2022;Tolston et al, 2018), galvanic skin response (Akash et al, 2018;Chen et al, 2015), interventions (Tenhundfeld et al, 2019(Tenhundfeld et al, , 2020, monitoring behaviors (Bahner et al, 2008;Bailey & Scerbo, 2007;Banks et al, 2018;Endsley, 2017), and eye-tracking (Hergeth et al, 2015;Lu & Sarter, 2019).…”

“…We have therefore run a study in which physiological, behavioral, and subjective self-report measures are used, to assess trust of participants interacting with an autonomous golf-cart. Some of the techniques we used have been routinely relied upon for trust assessment, such as subjective self-reports (Brown & Galster, 2004; Foroughi et al, 2021; Seet et al, 2022; Wojton et al, 2020), heart rate (Mitkidis et al, 2015; Tolston et al, 2018), and monitoring behaviors (Bailey & Scerbo, 2007; Ferraro et al, 2018; Tenhundfeld et al, 2019). However, we also selected measures that have been linked to trust through being measures of stress but have not had the same breadth of use in empirical studies, such as heart-rate variability (HRV) (Petersen et al, 2019) and displacement behaviors (Fratczak et al, 2021).…”

Assessment of Trust in Automation in the “Real World”: Requirements for New Trust in Automation Measurement Techniques for Use by Practitioners

“…There are various techniques used for characterizing the dynamics of the system, including attractor reconstruction (Abarbanel, 2012), stability analysis (Rosenstein et al, 1993), recurrence quantification analysis (Webber & Zbilut, 1994), and long-range correlation (Hurst, 1951). Some studies have already examined trust calibration by characterizing behavioral (Demir et al, 2021; Grimm et al, 2018) and physiological interaction dynamics (de Visser et al, 2018a; Mitkidis et al, 2015; Tolston et al, 2018; M. Wang et al, 2018), sometimes relying on the application of layered dynamics (Gorman et al, 2019; Grimm et al, 2018), and joint recurrence quantification (Demir et al, 2019) analyses methods. Much of the inspiration for these dynamic models comes from the work of John Gottman (e.g., Gottman, 2011; Gottman et al, 2005).…”