Trust in Automated Agents is Modulated by the Combined Influence of Agent and Task Type

Abstract

Trust in automation is an important topic in the field of human factors and has a substantial impact on both attitudes towards and performance with automated systems. One variable that has been shown to influence trust is the degree of human-likeness that is displayed by the automated system with the main finding being that increased human-like appearance leads to increased ratings of trust. In the current study, we investigate whether humanness unanimously leads to higher trust or whether the degree to which an agent is trusted depends on context variables (i.e., task type). For that purpose, we created a task with a social (i.e., judging emotional states from the eye region) and an analytical component (i.e., mathematical task) and measured how strongly participants complied to human, avatar or computer agents when performing the social versus the analytical version with them. We hypothesized that human-like agents are trusted more on social tasks, while machine-like agents are trusted more on analytical tasks. In line with our hypothesis, the results show that, human agents are in general not trusted more than automated agents but that the degree to which an agent is trusted depends on the anticipated expertise of an agent for a given task. The findings suggest that when designing automated systems that are supposed to interact with humans, the degree of humanness of the agent needs to match the degree to which a task requires social skills.

Good advice is beyond all price, but what if it comes from a machine?

“…the tendency to overly rely on initial information) which has been shown to reduce trust in received advice during decision-making and might explain why advisor type did not modulate post-decision trust (Goodyear et al, 2016;Madhavan & Wiegmann, 2005). In line with this hypothesis, a previous study conducted by Smith et al (2016) involving the same agents showed that when human, robot, and computer advisors were randomly assigned to either the social or analytical task, trust in the agents' advice showed the expected advisor by task type interaction, with stronger compliance with the human advisor on the social task and the machine agents on the analytical task. This indicates that allowing individuals to select their preferred advisor before receiving specific advice from them produces a different trust outcome than when agents are assigned without participant control.…”

“…Lost for Words; Golan et al, 2006). The analytical task was adapted from Smith, Allaham & Wiese (2016) and required participants to solve standard addition and subtraction operations.…”

“…Trust was measured in terms of compliance, which was calculated as the percentage of time participants changed their initial response to match the advice given by the chosen agent (Fox, 1996;Lee & Morray, 1992;Smith, Allaham, & Wiese, 2016). Due to the nature of the experiment, participants did not choose agents at equivalent rates.…”

“…It was found that trust in a machine advisor is modulated by general factors like its perceived expertise (Muir & Morray, 1996;Muir & Morray, 1994;Liu, 2010), but also factors that relate to its similarity to humans such as its human-like appearance (Smith, Allaham, & Wiese, 2016;Madhavan & Wiegmann, 2007) or sensitivity to the advisee's emotional states (Picard, 1997;Goetz et al, 2003). In particular, it was shown that agents with a humanlike physique receive higher trust ratings than mechanistic looking agents (Smith et al, 2016;Madhavan & Wiegmann, 2007), and that participants comply more with instructions given by human-like agents as opposed to machine-like agents (Goetz et al, 2003). Interestingly, however, when different agent features, such as the capacity to feel sensations such as hunger (i.e., experience) and the capacity to plan (i.e., agency), communicated inconsistent levels of humanness (e.g., agent seemed able to plan, but unable to feel), participants had negative attitudes towards the advisor and showed low levels of trust (although at least one of the features was high in humanness; Gray & Wegner, 2012).…”

A Little Anthropomorphism Goes a Long Way

“…The lack of oxytocin differences for the human agent may follow in this tradition, as overall trust levels were lower for human agents. Indeed, a previous study using the same pattern-matching task found that participants held a strong bias against collaborating with the human (de Visser et al, 2016), an effect that seems to be driven by initially lower expectations regarding the human agent’s ability to complete the pattern-matching task accurately (Smith, Allaham, & Wiese, 2016). Lower levels of trust and compliance observed in the current study suggest that the human agent was again seen as untrustworthy: administering oxytocin may not be able to override such a bias consistent with previous research (Mikolajczak et al, 2010).…”