Tobias M. Peters

Anthropomorphizing artificial intelligence (AI) -- that is, attributing human characteristics to it -- is prevalent in public discourse. Using human-like voices or appearances is a prominent example. Similarly, the language used to describe AI applications is anthropomorphic. AI is "fed" with data, it "learns", "acts", and "decides". However, little is known about the effects of using such anthropomorphisms when describing AI. Previous research has shown that anthropomorphic appearance and design increase trust among people interacting with AI applications. To test whether this translates to anthropomorphism used in descriptions of AI, we conducted two pre-registered experiments comparing two conditions in which an AI application is either linguistically anthropomorphized or not. In the first experiment, participants read a fictitious newspaper article about the usage of an AI application in radiology. In the second experiment, participants completed a classification task using AI. Using a response-elicitation approach, participants had to judge their trust and distrust towards the AI system, and reliance was compared in the second experiment. We found no significant differences in trust, distrust, or reliance between anthropomorphic and non-anthropomorphic conditions. Subgroup analyses accounting for domain expertise and general attitudes towards AI yielded similar null results. These results suggest that anthropomorphic descriptions alone are insufficient to shift trust-related concepts (trust, distrust, and reliance). Further, the results indicate boundaries for metaphorical framing effects in general and, specifically, for trust.

When humans interact with artificial intelligence (AI), one desideratum is appropriate trust. Typically, appropriate trust encompasses that humans trust AI except when they either explicitly notice AI errors or when they are suspicious that errors could be present. So far, appropriate trust or related notions have been mainly investigated by assessing trust and reliance. In this contribution, we argue that these assessments are not sufficient to measure the complex aim of appropriate trust and the related notion of healthy distrust. We introduce and test the perspective of visual attention as an additional indicator for appropriate trust and draw conceptual connections to the notion of healthy distrust.To test the validity of our conceptualization, we formalize visual attention using TVA and measure its properties potentially relevant to appropriate trust and healthy distrust in an image classification task. Thereby, we investigate participants’ attentional capacity and attentional weight towards correct and incorrect classifications. We observe that misclassifications reduce attentional capacity compared to correct classifications. However, our results do not indicate that this reduction is beneficial for a subsequent judgment of the classifications. The attentional weighting is not affected by the classifications’ correctness but by the difficulty of categorizing the stimuli themselves. These results, their implications, and the limited potential for using visual attention as an indicator of appropriate trust and healthy distrust are discussed.

Introduction: When humans interact with artificial intelligence (AI), one desideratum is appropriate trust. Typically, appropriate trust encompasses that humans trust AI except for instances in which they either explicitly notice AI errors or are suspicious that errors could be present. So far, appropriate trust or related notions have mainly been investigated by assessing trust and reliance. In this contribution, we argue that these assessments are insufficient to measure the complex aim of appropriate trust and the related notion of healthy distrust. We introduce and test the perspective of covert visual attention as an additional indicator for appropriate trust and draw conceptual connections to the notion of healthy distrust. Methods: To test the validity of our conceptualization, we formalize visual attention using the Theory of Visual Attention and measure its properties that are potentially relevant to appropriate trust and healthy distrust in an image classification task. Based on temporal-order judgment performance, we estimate participants' attentional capacity and attentional weight toward correct and incorrect mock-up AI classifications. We observe that misclassifications reduce attentional capacity compared to correct classifications. However, our results do not indicate that this reduction is beneficial for a subsequent judgment of the classifications. Results and discussion: The attentional weighting is not affected by the classifications' correctness but by the difficulty of categorizing the stimuli themselves. We discuss these results, their implications, and the limited potential for using visual attention as an indicator of appropriate trust and healthy distrust.