Abstract
Trust is fundamental to creating a lasting relationship with another human being. In our daily lives, we encounter several situations where we place trust on other humans such as bus drivers, colleagues, and even strangers. Trust in human-human teams is initially founded on the predictability of the trustee, and as the relationship between the trustor and the trustee progresses, dependability or integrity replaces predictability as the basis of trust (Hoff & Bashir, 2015). As artificial intelligence (AI) gets smarter and smarter, it is becoming an integral part of human lives. For example, AI is being used for scheduling appointments, powering smart homes, recognizing people’s faces in photos, making health diagnoses, providing investment advice, and several other things. As with a human-human team, trust is a necessary ingredient for human-AI partnerships. However, trust in human-human teams progresses in the reverse order from human-human teams (e.g., Hoff & Bashir, 2015). This is because humans initially assume that AI is near perfect. Therefore, in the initial stages, faith forms the essential constituent of trust, and as the number of exchanges between the human and the AI increases, faith is replaced by dependability and predictability.
Trust in AI is considered a two-dimensional construct comprising trust and distrust, where trust is associated with feelings of calmness and security and distrust involves fear and worry (Lyons, Stokes, Eschleman, Alarcon, & Barelka, 2011). Unarguably, trust is a complex social process with a variety of factors determining the extent to which humans trust AI agents (Hoff & Bashir, 2015; Lee & See, 2004). Specifically, in order to calibrate the right level of trust in AI, consider dispositional (i.e., user characteristics such as age, culture, gender, and personality), internal (e.g., user characteristics such as workload, mood, self-confidence, and working memory capacity), environmental (e.g., task difficulty, perceived risks and benefits, organizational setting), and learned factors (e.g., reputation of the AI, reliability, consistency, type and timing of errors made by the AI, and users’ experience with similar agents).
Design factors (e.g., appearance, ease of use, communication style, and transparency of the AI) also affect perceptions of trust. For example, anthrophomorphizing is an effective way of establishing a long-term social bond between humans and AI agents, which is driven by the neurotransmitter and hormone, oxytocin (e.g., de Visser et al., 2017). Further, anthropomorphic agents also resist breakdown in trust compared to their counterpart non-anthropomorphic agents, presumably because anthropomorphic agents remind users of humans who are forgiven more easily for being imperfect in comparison to machine-like agents. Interestingly, there is also evidence that humans tend to disclose more information to AI therapists than human therapists. Transparency of the AI also helps to calibrate the right level of trust by enabling users to develop accurate mental models of the AI underpinnings (e.g., Balfe, Sharples, & Wilson, 2018).