Cognitive Model of Trust Dynamics Predicts Human Behavior within and between Two Games of Strategic Interaction with Computerized Confederate Agents

The Importance of Being Consistent: Attribution of Mental States in Strategic Human-Robot Interactions

This article investigates the attribution of mental state (AMS) to an anthropomorphic robot by humans in a strategic interaction. We conducted an experiment in which human subjects are paired with either a human or an anthropomorphic robot to play an iterated Prisoner's Dilemma game, and we tested whether AMS is dependent on the robot "consistency," that is, the correspondence between the robot's verbal reaction and its behavior after a nonoptimal social outcome of the game is obtained. We find that human partners are attributed a higher mental state level than robotic partners, regardless of the partner's consistency between words and actions. Conversely, the level of AMS assigned to the robot is significantly higher when the robot is consistent in its words and actions. This finding is robust to the inclusion of psychological factors such as risk attitude and trust, and it holds regardless of subjects' initial beliefs about the adaptability of the robot. Finally, we find that when the robot apologizes for its behavior and defects in the following stage, the epistemic component of the AMS significantly increases.

Trust Miscalibration Is Sometimes Necessary: An Empirical Study and a Computational Model

The literature on trust seems to have reached a consensus that appropriately calibrated trust in humans or machines is highly desirable; miscalibrated (i.e., over- or under-) trust has been thought to only have negative consequences (i.e., over-reliance or under-utilization). While not invalidating the general idea of trust calibration, a published computational cognitive model of trust in strategic interaction predicts that some local and temporary violations of the trust calibration principle are critical for sustained success in strategic situations characterized by interdependence and uncertainty (e.g., trust game, prisoner’s dilemma, and Hawk-dove). This paper presents empirical and computational modeling work aimed at testing the predictions of under- and over-trust in an extension of the trust game, the multi-arm trust game, that captures some important characteristics of real-world interpersonal and human-machine interactions, such as the ability to choose when and with whom to interact among multiple agents. As predicted by our previous model, we found that, under conditions of increased trust necessity, participants actively reconstructed their trust-investment portfolios by discounting their trust in their previously trusted counterparts and attempting to develop trust with the counterparts that they previously distrusted. We argue that studying these exceptions of the principle of trust calibration might be critical for understanding long-term trust calibration in dynamic environments.

Investigating the effect of trustworthiness on instruction-based reflexivity

Unlike other species, humans are capable of rapidly learning new behavior from a single instruction. While previous research focused on the cognitive processes underlying the rapid, automatic implementation of instructions, the fundamentally social nature of instruction following has remained largely unexplored. Here, we investigated whether instructor trustworthiness modulates instruction implementation using both explicit and reflexive measures. In a first preregistered study, we validated a new paradigm to manipulate the perceived trustworthiness of two different virtual characters and showed that such a manipulation reliably induced implicit associations between the virtual characters and trustworthiness attributes. Moreover, we show that trustworthy instructors are followed more frequently and faster. In two additional preregistered experiments, we tested if trustworthiness towards the instructor influenced the cognitive processes underlying instruction implementation. While we show that verbally conveyed instructions led to automatic instruction implementation, this effect was not modulated by the trustworthiness of the instructor. Thus, we succeeded to design and validate a novel trustworthiness manipulation (Experiment 1) and to create a social variant of the instruction-based reflexivity paradigm (Experiments 2 and 3). However, this instruction-based reflexivity effect was not modulated by the instructors' trustworthiness.

Toward a Unified Theory of Learned Trust in Interpersonal and Human-Machine Interactions

A proposal for a unified theory of learned trust implemented in a cognitive architecture is presented. The theory is instantiated as a computational cognitive model of learned trust that integrates several seemingly unrelated categories of findings from the literature on interpersonal and human-machine interactions and makes unintuitive predictions for future studies. The model relies on a combination of learning mechanisms to explain a variety of phenomena such as trust asymmetry, the higher impact of early trust breaches, the black-hat/white-hat effect, the correlation between trust and cognitive ability, and the higher resilience of interpersonal as compared to human-machine trust. In addition, the model predicts that trust decays in the absence of evidence of trustworthiness or untrustworthiness. The implications of the model for the advancement of the theory on trust are discussed. Specifically, this work suggests two more trust antecedents on the trustor's side: perceived trust necessity and cognitive ability to detect cues of trustworthiness.