Abstract: As artificial intelligence expert systems (AIES) are increasingly used in domains such as medicine, psychotherapy, law, finance, and education, it remains important to understand what shapes their perceived trustworthiness relative to human experts. A central limitation of prior research is that findings are often difficult to generalize across domains because domain and stakes are rarely varied together within the same design. We address this issue in a preregistered online vignette experiment in which UK participants ( N = 898) evaluated advice from an AIES and a human expert under low- and high-stakes conditions in one of five domains. Across all domains and both stake levels, human experts were perceived as more trustworthy than AIES, whereas AIES were perceived as riskier. For trustworthiness, between-domain effects were significant but small on average, mainly affecting the size of the human–AIES difference, which was largest in medicine and psychotherapy and smallest in education. For perceived risk, stakes played a larger role than domain, with high-stakes situations increasing perceived risk across domains. Importantly, trustworthiness and perceived risk did not change in parallel across conditions: Domain differences were comparatively small for trustworthiness, whereas stakes played a larger role in perceived risk. Overall, the findings suggest that cross-domain differences are smaller than often assumed once stakes are controlled, and that stakes matter more than domain for understanding perceived risk when people evaluate AIES advice.
Ehrhardt et al. (Fri,) studied this question.